FR2843469A1 - Visual perception method for object characterization and recognition through the analysis of mono- and multi-dimensional parameters in multi-class computing units and histogram processing - Google Patents

Abstract

Method for perceiving an object in space using spatial resolution variation to characterize and recognize an object. Accordingly the object is represented as a series of sequences and sub-sequences of a digital signal and the inventive method is repeated a number of times. The method involves formation of a Gaussian time variation in relation to the object spatial resolution within the signal, followed by filtering and parsing to obtain prioritized details. The invention also relates to a corresponding active visual perception device, and multiclass mono- and multi-dimensional histogram processing and computer modules that can be dynamically recruited.

Description

The present invention relates to a method of

  operation of a multiclass functional unit and a multiclass functional unit for calculation and histogram processing.

  The multi-class functional unit can be implemented in a calculation block and histogram processing unit for analyzing a

  or several characteristic parameters of spatiotemporal space objects that are represented by sequences and subsequences of data evolving over time. The invention can be applied to the analysis of images as video data for the purpose of identifying and locating objects of a scene.

  The analysis allows the identification of an object by its shape and / or its dimension and / or its orientation and / or its respective position with respect to the scene and / or other objects of the scene. An application of the invention to active visual perception is more particularly detailed in this document. The invention can also be implemented with other types of data that can be represented in the form of sequences and

  sub-sequences of data such as for example sounds.

  Methods and devices for locating objects in an image are already known. In some of them it is proposed the statistical analysis of the points or pixels of a digital video signal coming from an observation system for the realization

  effective devices that can operate in real time.

  More recently, it has been proposed to implement these devices 25 by the association of information processing units of the same nature each addressing a particular parameter extracted from the video signal to analyze it. This is the case of the patent applications of Patrick PIRIM, inventor WO-98/05002, FR-01/02539 filed on February 23, 2001, WO00 / 11609, WO-01/63557 or the article "The mechanism of the vision integrates on a chip "in Electronics, June 2000 N 104 in which the implementation of blocks or units or modules is proposed, these terms being equivalent here, calculation and histogram processing constituting true space neurons electronic time, said STN, each analyzing a parameter, said parameter being processed by a

  function (fog) to individually produce an output value.

  These output values, all together, form an available feedback on a bus for use in the blocks during the analysis. At the same time, each of these calculation and histogram processing units constitutes and updates an analysis output register providing statistical information on the corresponding parameter. The choice of the parameter analyzed by each calculation block and histogram processing, the contents of the analysis output register and the function (fog) that it performs, are determined by software executed in an application programming interface. API (Application Program Interface). In a calculation block and histogram processing as described in WO01 / 63557, for a given parameter, it is determined from the histogram calculated and stored in a memory, the maximum RMAX of the histogram, the position of said maximum POSRMAX, a number of NBPTS points of the histogram. Classification boundaries are also determined which make it possible to delimit an area of interest for the parameter and it has been proposed to take as a criterion for the determination of the limits a ratio of the maximum of the histogram, for example RMAX / 2, and to obtain the terminals by scanning the data from the memory from the origin in search of the corresponding zone boundaries

to the criterion.

  An application of the STN blocks is more particularly detailed in the application FR-01/02539 where it has been proposed to hierarchically break down the object to be identified according to its properties, which makes it possible, for example, to first determine the general contour. of an object moving relative to a relatively stable background 30, then search within this contour of the characteristic elements by their hue, their color, their relative position ... Such an approach allows the rapid development of multiple applications involving the tracking of an object. These applications can be developed either from an earlier formalization that has revealed the significant characteristics of the object, or through a learning function by examining a scene in which the object in question is present, the device allowing itself to extract

  characteristic parameters of the object.

  It has therefore been proposed in this application FR01 / 02539 a method of locating a shape in a space represented by pixels forming together a space i, j multidimensional, evolving over time, and represented at a succession of moments, said data each associated with a temporal parameter A, B, ... being in the form of DATA (A), DATA (B), ... signals consisting of a sequence Aijt, Bijt,

  . binary numbers of n bits associated with synchronization signals making it possible to define the moment of the space and the position i, j in this space, to which the signals Aijt, Bijt, ... received at a given instant. In this application, it is indicated that: a) an area of interest of the space is identified according to a statistical criterion applied to a temporal parameter, b) the main zone thus identified is inhibited, c) the steps a) and b) so as to locate other areas of interest within a zone of uninhibited space, d) the process is stopped when a remaining, uninhibited area of the space no longer produces a zone of interest corresponding to the statistical criterion, in other words, when, in the area of interest, the number of points is too small (less than a threshold), e) it is incremented by consecutive valid frame, a counter for each zone of interest thus identified, the centroid of its cloud of points, .. DTD: f) for each area of interest thus identified, the center of gravity of its cloud of points is recovered.

  In this application it has also been shown that it is advantageous to implement two subsets of calculation blocks and histogram processing receiving the signals and each producing a classification value, the first subset receiving a signal carrying a signal. a first temporal parameter and the second subset receiving two spatial signals, the classification value of the first subset validating a group of space points processed by the second subset, the number of said points being ni, the value the second subassembly validating the parameter values processed by the first subset, the two subsets jointly producing a binary signal representing an area of interest and a signal representing the value of the temporal parameter in that area. In various applications it has also been shown that it is advantageous to associate with this device a third subset receiving a signal carrying a second temporal parameter, this third subset having a function similar to the first and replacing it. when validating points of space whose number

is n2, n2 being greater than n1.

  In these applications it has also been proposed to implement several subsets receiving spatial signals enabling successively to validate several groups of points in the space and finally to produce a device which comprises a set of calculation blocks and histogram processing. 20 controlled by API software and connected to each other by a bus of

  data and by a feedback bus.

  If such a computational block and histogram processing device makes it possible to reliably identify and locate an object of a scene, it has some limitations due to the method of determining the classification boundaries and the

  the type of the parameter being analyzed. It is therefore an object of the invention to provide an improved method and apparatus employing calculation blocks and histogram processing which allow greater efficiency and flexibility in histogram analysis.

  Such an approach is related to the fact that we realized that it was more efficient to determine on a histogram in a given block several vertices for classification, each of the vertices

  corresponding to a particular class.

  To this end, a method of operating a multiclass functional unit for calculating and histogram processing for analyzing a parameter carried by a digital DATA (A) input signal in the form of a sequence Aij is proposed. t, A'ij ... t, 5 A "ij ... t ... of binary numbers associated with synchronization signals making it possible to define a moment t of space and a position i, j ... in this space, the histogram being able to comprise several vertices and the processing consisting in producing a set of results of characterization of the histogram lo comprising at least the data pair maximum amplitude of the histogram and position of said maximum amplitude

  corresponding to the maximum peak of the histogram.

  According to the invention, the histogram of the parameter is calculated as a function of a validation signal (VALIDATION) during a calculation cycle, and which is determined during said calculation cycle.

  a set of pairs of amplitude RMAXi and position POSRMAXI, the pairs RMAXI and POSRMAXI being automatically classified in descending order of vertex and stored in a functional memory with automatic material sorting of the multiclass functional unit during said calculation cycle.

  The invention can also be implemented with the following characteristics, possibly combined according to all the technically feasible possibilities: a set of results is furthermore produced in the form of k classification signals (C11 ... Clk) with k greater than or equal to 1, each classification signal corresponding to a class related to a vertex of the histogram, each of the classes of the classification signal is produced in relation to a set of parameter values for which the amplitude is of the histogram is greater than a threshold criterion, - the threshold criterion is a function of the value of the amplitude of the considered vertex of the histogram, - the threshold criterion is selectable, - a functional unit is implemented multiclass which comprises an application programming interface (API), a memory M0 storing the ordinate amplitudes RMAXi of the vertices, a memory M1 of addresses amplitude ordinates POSRMAXi of vertices, memories M0 and M1 being grouped into a single functional table of amplitude and memory position pairs 5 (RMAXo / POSRMAXo; RMAX1 / POSRMAX1; RMAX2 / POSRMAX2, RMAXI / POSRMAXI ...) vertices of the histogram in descending order of amplitude, the functional table performing an automatic hardware sorting of classes during the computation cycle, the multiclass functional unit further comprising a memory M2 lo making it possible to store the class order number, a class threshold memory M3 and a class number register RC, - the API makes it possible to carry out: - an initialization cycle setting the memories to zero M0, M1, M2, M3 and the register, - a calculation cycle for automatically determining and classifying in the functional table the pairs of values, - a class update cycle, the update cycle comprising: (B a step of searching classes with labeling of classes in the memory M2 and storing the corresponding thresholds in the memory M3, the number of classes thus determined being stored in the register RC, (C) - a step of validating the classes of l a memory M2 25 by comparing the value of the memory M0 to the address of the considered class of M2 with the threshold of M3 corresponding to the class considered, - the API is a programmable sequencer of the microprocessor type with memory or microcontroller, - In the case of a functional table is implemented a set of sorting and storage units B0 ... Bn with memory registers, comparators and associated logic circuits for automated classification and storage of amplitudes and positions vertices of the histogram, - the number of points NBPTSi of the histogram corresponding to said class is determined and memorized further for each class, in a memory M4, - for each class, it is furthermore determined and memorized, in a memory M5, the average POSMOYI position of said parameter in said class; furthermore, a barycentric unit is used for producing a barycenter output signal with a p first binary state when the parameter corresponds to the average position o10 POSMOYi and with a second binary state otherwise for the class considered, - it implements a selection means for selecting for a given class a threshold criterion based at least one of the following values: the value of the amplitude of the top of the histogram, a threshold value THRESH supplied to the unit, a number of NBPTS points of the histogram, each of the classes of the a classification signal being produced in relation to a set of parameter values for which the amplitude of the histogram is greater than said threshold criterion, - for a given class the criterion is selected from RMAX / 2,

THRESHOLD, NBPTS / THRESHOLD,

  the parameter analyzed by the multiclass functional unit is complex and is obtained by combining at least two elementary parameters, each of the binary numbers of the input signal DATA (A1, A2, A3 ... Ap) support of the complex parameter (A1A2A3 ... Ap) comprising P fields each corresponding to an elementary parameter A1, A2, A3,..., Ap, - the multiclass functional unit is implemented in a module 30 for computing and processing the histogram, said STN, the output signals forming a retro-annotation being sent on a retro-annotation bus and the validation signal (VALIDATION) involving the retro-annotation; - the multiclass functional unit receives at least one parameter, a signal of validation (VALIDATION), linear combination of feedback signals in the case of an STN module, and sequencing signals (INIT, CALCUL, END, CLOCK), - the multiclass functional unit returns at least one set of signals corresponds each to a class (ClI ... Clk) on the feedback bus in the case of an STN module, - in an object recognition system comprising a set of calculation modules and histogram processing, it is possible to at least two multi-class functional unit STN modules, a first module operating in a time domain TD, determining at least one class and recruiting for said class at least one second module operating in a spatial domain SD, the module operating in the time domain TD receives a speed parameter MVT or of color L / TIS, the method is implemented in an object recognition system comprising a set of STN modules for calculation and histogram processing by sectorization with at least a multi-class functional unit module, the modules defining zones and centroids, and in that a zone 20 is divided into several angular sectors centered on the corresponding center of gravity. the area and looking in which sector among the sectors, a new barycentre appears and divide the sector into several sub-sectors, this process being able to continue in order to continue to progressively sharpen the sectorisation, - two levels of sectoring, a first dividing an initial zone into sectors and a second dividing one of the sectors having a new centroid into sub-sectors, - at the end of the sectorization at least one angle and one module are determined, the angle being given with reference to the straight line joining the two centers of gravity and the module corresponding to the distance along said line between said two centers of gravity, sectorization is carried out in four sequences, the method is implemented in at least one STN module with 35 at least two pointer units poc pp of input axes for rotation of reference axes of at least two cartesian coordinates of input parameters, the Moreover, the modules determine the center of gravity for the input parameters, and in the system a first space Zi having a BarZ center of gravity is determined by a combination of a monolinear module processing a first parameter and a second bilinear module processing the coordinates, a second association determining within said first space Zi the appearance of a second barycenter BarZ1 + 1, said first space being split into distinct regularly distributed angular sectors (Zr1o, Zr1l, Zri2 ...), each sectors being treated by a bilinear sector module receiving ZI, BarZi and also the signal of the second barycenter BarZ, + 1, the bilinear sector module corresponding to the second barycenter BarZ1 + 1 being related to a set of bilinear modules of sector of subsequent rank to split the sector having the second barycenter BarZi + 1 in separate regular angular sectors (Zra, Zrb, Zr0 ...) in order to progressively sharpen the sectorization and furthermore determine at least one angle α of reference axes 20 substantially perpendicular to the direction of the straight line uniting the two BarZi and BarZi, + 1 and one module barycentres. pj distance between the two barycentres BarZ1 and BarZ1, + 1 along said straight line, the reference according to the angle a to obtain a translational invariance, - one calculates pj, a, Ca, - object can be observed at different distances and is implemented in the system furthermore at least one size invariance unit, said size invariance unit receiving at least an input, on the one hand, a value of the logarithm of a distance LD 30 between a reference point and at least one point of the object and, on the other hand, the module pj of distance between two centroids BarZ1 and BarZi + 1, said unit determining at least a substantially constant projection value Cp 'and corresponding to a rotation angle 0 In the reference frame p and LD, ALD, 0, Cp 'are calculated. In the unit of size invariance, distance control means are used which make it possible to use a measurement of the distance. external distance or the internal determination of the distance, the object is observed at different distances by physical displacement or by zooming effect, the object can be observed at different angles and implemented in the system. in addition to at least one rotation correction unit (900), said rotation correction unit 10 for correcting the reference axis angle α with respect to a pair of values a, p of angle and modulus previously determined, the rotational correction unit (900) further makes it possible to determine the angle of rotation Au; the method is implemented with at least one parameter marker transformation unit by rotation of angle θ, the mark being at least on two dimensions of 2, the rotation in the case of a two-dimensional landmark for pixel coordinate parameters X, Y, corresponding to the following matrix operation: FX1 Fcos 0 sin 0 1 [xl

= X

  sin 0 -cos 0 LYJ - the angle 0 (or a) is chosen so that at least the projection of one of the parameters on the corresponding axis (pO) (pox) has a reduced rate of change after rotation, - the parameters are two in number and are chosen from the pairs X, Y of pixel coordinates or LogD, p of logarithm of distance of one pixel with respect to a point of reference and of angle with respect to said point and with a reference line, - is determined over time several barycentres of zones 30 appearing successively relative to a first of an initial zone and stored the center of gravity coordinates in relation to a label in the form of recognition data, - it realizes an analysis tree linking the different centers of gravity according to their order of appearance, to determine whether or not a new observed object corresponds to an object with label previously stored, the recognition data of the new object is determined and compared to previously recognized label recognition data, - the label recognition data is further associated with an angle of rotation Ao and a mean distance LD 'by means for determining said angle of rotation Aa and said average distance LD', in addition, the label recognition data is associated with a dominant color C by means making it possible to determine said color; recognition data of a first label is associated with those of at least one second label in order to form a label; new label corresponding to a higher level of recognition, - the label recognition data is analyzed by a calculation and histogram processing module capable of determining and storing a set of data of

categorization of said labels.

  The invention also relates to a device which is a multi-class functional unit for operating according to one or more of the preceding functional characteristics relating to the method. This invention therefore relates to a multiclass functional unit for calculating and histogram processing for analyzing a parameter carried by a digital DATA (A) input signal in the form of a sequence Aij ... t, A'ij. ..t, A "ij ... t ... of binary numbers associated with synchronization signals making it possible to define a moment t of space and a position i, j ... in this space, the histogram being able to having a plurality of vertices and the processing of producing a set of histogram characterization results comprising at least the maximum amplitude histogram data pair and the position of said maximum amplitude corresponding to the maximum peak of the histogram. device invention, the unit comprises means 35 for calculating the histogram of the parameter as a function of a validation signal (VALIDATION) during a calculation cycle, and for determining during said computation cycle a set couples of amplitude R MAXI and position POSRMAXi, the pairs RMAXj and POSRMAXI being automatically classified in order of 5 descending vertex and memorized in a functional memory with automatic material sorting of the multiclass functional unit

during said calculation cycle.

  The unit can also be implemented with the following characteristics, possibly combined according to all the technically feasible possibilities: the means furthermore make it possible to output a set of results in the form of k classification signals (C1 ... Clk) with k greater than or equal to 1, each classification signal corresponding to a class related to a top of the histogram, - each of the classes of the classification signal is produced in relation to a set of parameter values for which the amplitude of the histogram is greater than a threshold criterion, the threshold criterion is a function of the value of the amplitude of the considered vertex of the histogram, the threshold criterion is selectable, the unit comprises an application programming interface (API), a memory M0 storing the ordinate amplitudes RMAXI of the vertices, a memory M1 of addresses ordered in amplitude 25 PO SRMAX1 of vertices, memories M0 and M1 being grouped into a single functional table of amplitude and memory position pairs (RMAXo / POSRMAXo; RMAX1 / POSRMAX1; RMAX2 / POSRMAX2, ... RMAXi / POSRMAXi ...) vertices of the histogram in descending order of amplitude, the functional table performing an automatic material sorting of the classes during the computation cycle, the multi-class functional unit comprising furthermore, a memory M2 making it possible to store the sequence number of the class, a class threshold memory M3 and a class number register RC, the API makes it possible to carry out: an initialization cycle setting to zero the memories MO, M1, M2, M3 and the register, - a calculation cycle for automatically determining and classifying in the functional table the pairs of values, - a class update cycle, the updating cycle. comprising: (B) - a step of searching classes with class labeling in the memory M2 and storing the corresponding thresholds lo in the memory M3, the number of classes thus determined being stored in the register RC, (C) - a step validation of clas its memory M2 by comparing the value of the memory M0 to the address of the considered class of M2 with the threshold of M3 corresponding to the class considered, - the API is a programmable sequencer of the microprocessor type with memory or microcontroller, - in the case of a functional table is implemented a set of sorting and storage units B0 ... Bn with memory registers 20, comparators and associated logic circuits for automated classification and storage of magnitudes and positions of the vertices of the histogram, the unit makes it possible to determine and memorize further for each class, in a memory M4, the number of points NBPTS 1 of the histogram corresponding to said class, the unit allows to determine and memorize further for each class, in a memory M5, the average position POSMOY1 of said parameter in said class, - the unit further comprises a barycentric unit for 30 pr deriving a barycenter output signal with a first binary state when the parameter corresponds to the average position POSMOYi and with a second binary state otherwise for

the class considered.

  The unit comprises a selection means making it possible to select for a given class a threshold criterion as a function of at least one of the following values: the value of the amplitude of the top of the histogram; a threshold value; THRESHOLD provided to the unit, - a number of NBPTS points of the histogram, each of the classes of the classification signal being produced in relation to a set of parameter values for which the amplitude of the histogram is greater than said criterion of threshold, - for a given class, the criterion is selected from RMAX / 2,

THRESHOLD, NBPTS / THRESHOLD,

  the parameter analyzed is complex and obtained by combining at least two elementary parameters, each of the binary numbers of the input signal DATA (Al, A2, A3, Ap) supporting the complex parameter ( A1A2A3 ... Ap) having P fields

  each corresponding to an elementary parameter A1, A2, A3, ...

  Ap, the unit is in a calculation module and histogram processing, said STN, the output signals forming a retroannotation being sent on a retroannotation bus and the validation signal (VALIDATION) involving the retroannotation, - The multiclass functional unit receives at least one parameter, a validation signal (VALIDATION), a linear combination of feedback signals in the case of an STN module, and sequencing signals (INIT, CALCUL, END, CLOCK), The multiclass functional unit returns at least one set of signals each corresponding to a class (Cl1 ... Clk) on the feedback bus in the case of an STN module, the unit is in a recognition system of object having a set of calculation modules and histogram processing

  comprising at least two multi-class functional unit STN modules, a first module operating in a time domain TD, determining at least one class and recruiting for said class at least one second module operating in a spatial domain SD.

  the module operating in the time domain TD receives a speed parameter MVT or of color L / T / S, the unit is in an object recognition system comprising a set of STN modules for computing and processing of histogram by sectorization with at least one multi-class functional unit module, and the modules make it possible to determine zones and centroids and to divide a determined zone into several angular sectors centered on the corresponding barycenter of the zone and to search in which sector among: o the sectors, a new center of gravity appears and to divide the sector into several sub-sectors, this process being able to continue in order to continue progressively refining the sectorization, - two levels of sectorization are carried out, one dividing an initial zone into sectors and one second, dividing one of the sectors with a new centroid into sub-sectors, - at the end of the sectorization we determine at least one year and a module, the angle being given with reference to the straight line joining the two centers of gravity and the module corresponding to the distance along said line between said two centers of gravity, the division is carried out in four sequences, the unit is in the system comprising at least one STN module with at least two orientation units of the input axes for rotation of reference axes of at least two Cartesian coordinates of input parameters, the module (s) further determining the center of gravity for the input parameters, and in the system determining a first space Zi having a Barij center by a combination of a monolinear module processing a first parameter and a second bilinear module 30 dealing with the coordinates, a second association determining within said first space Zi the appearance of a second center of gravity BarZi + 1, said first space being divided into distinct angular sectors; (Zr10, Zr1l, Zri2 ...), each of the sectors being treated by a bilinear sector module receiving Zi, BarZi and also the signal of the second barycenter BarZ1 + 1, the bilinear sector module corresponding to the second barycenter BarZi + l being put in relation with a set of bilinear modules of subsequent rank sector making it possible to split the sector having the second barycenter 5 BarZl + in regularly distributed distinct angular sectors (Zra, Zrb, Zrc ...) in order to progressively refine the sectorization and further determine at least one angle α of reference axes substantially perpendicular to the direction of the straight line uniting the two centers BarZi and BarZi + 1 and a module pj distance between the two barycentres BarZ1 and BarZi + 1 along of said straight line, the reference according to the angle a allowing to obtain a translation invariance, - the values pj, a, Ca, are calculated, - the object can be observed at di stanzas and the system 15 further comprises at least one size invariance unit, said size invariance unit receiving at least one input, on the one hand, a value of the logarithm of a distance LD between a point of reference and at least one point of the object and, on the other hand, the module pj of distance between two centroids BarZi and BarZ, + 1, said unit 20 determining at least one projection value Cp 'substantially constant and corresponding to a 0 angle of rotation relative to the reference p and LD, - the values ALD, 0, Cp ', are calculated, - the size invariance unit comprises distance control means allowing the choice the use of external distance measurement or internal distance determination, the object is observed at different distances by physical displacement or zooming effect, the object can be viewed at different angles and the system includes in addition at least one unit rotation correction means, said rotation correction unit for correcting the reference axis angle α with respect to a previously determined pair of values a, p of angle and modulus, - the correction unit of rotation furthermore makes it possible to determine the angle of rotation Auc, the unit is associated with at least one parameter marker transformation unit by rotation of angle θ, the mark being at least on two parameter dimensions, the rotation in the case of a two-dimensional mark for pixel coordinate parameters X, Y, corresponding to the following matrix operation FX Fcos 0 sin 0 - X Sin 0 -cos 0] y - the angle O (or ca) is chosen so that at least the projection of one of the 1o parameters on the corresponding axis (pO) (poz) has a reduced rate of change after rotation, - the parameters are two in number and are chosen from the X, Y pairs of pixel coordinates or LogD, p of logarithm of d istance of a pixel relative to a reference point and angle 15 relative to said point and to a reference line, - in the system means can determine over time several centroids of areas appearing successively relative to a first of an initial zone and to memorize the centroid coordinates in relation with a label 20 in the form of recognition data, - an analysis tree is created which links the different centroids according to their order of appearance, - for whether or not a new observed object corresponds to a previously stored label object, the system makes it possible to determine the recognition data of the new object and compare it with previously stored label recognition data; recognition of the label are associated an Au rotation angle. and a mean distance LD 'by means for determining said angle of rotation Ac and said average distance LD'; furthermore, the label recognition data is associated with a dominant color C by means for determining said color; associating recognition data of a first label with those of at least one second label in order to form a new label corresponding to a higher level of recognition; - the system makes it possible to analyze the recognition data of 5 labels by one calculation module and histogram processing capable of determining and storing a set of

categorization of said labels.

  The invention as a method or device according to one or more of the above optionally combined features preferably operates on video image data. The space of the data in which the object is to be located is preferably a spatiotemporal space, that is to say that on the one hand it evolves over time and that on the other hand the parameters of the object can be represented at a given moment on a 15 (point of space) or two (surface) or three (volume or hue + saturation + light or other) or more dimensions as appropriate. The method and apparatus of the invention can be applied to active visual perception to characterize and recognize an object, especially for identification and

  location. This application will be more particularly detailed in the illustrative part of this description with the details of the invention presented in relation to the structure and operation of the STN blocks. The application therefore relates to a system for the perception, recognition and

  locating an object in its environment from a digital input signal constituted by a succession of successive sequences of views of the object in its environment and thus falling within the time domain, each of said sequences being constituted by a succession sub-sequences, each representative of locations arranged one after the other in said sequences and therefore falling within the spatial domain, characterized in that a temporal variation of the spatial resolution of the object in said input digital signal, the variation comprising a substantially Gaussian increase phase of the resolution from a reduced value to an optimum base value, differentiation is further carried out, with smoothing between two successive sequences of said substantially Gaussian increase in resolution, in order to obtain a digital derivative signal representative of the variability of the Gaussian difference between these two sequences when the difference, in absolute value, for each same spatial location of said derived signal, exceeds a threshold, and deduced, from said derived signal, lo by comparison in the successive sequences between the values of this derived signal corresponding to identical locations in the sub-sequences, of hierarchical details of the object, forming at least two histograms of said derived digital signal, of which at least one is relative to the numerical magnitude of said signal in the various locations, which provides information relating to the characteristic details of the object, and at least one other relates to the location of the locations in said signal, which provides location information

said details.

  The invention will now be described in more detail, but without any limitation, with reference to the accompanying drawings in which: FIGS. 1 and 2 show two embodiments of a device for collecting or knowing an object according to invention comprising: means for performing, during a period, a substantially Gaussian incremental increase in spatial resolution, consisting of an optical assembly for the first embodiment and an electronic filter for the second; a unit for performing a differentiation of Gaussians by spatio-temporal smoothing; and means for using the differentiation. FIG. 3 represents, by a curve, the desired variation of the spatial resolution, comprising, during a period, a

  incremental increase between r min and r max.

  FIGS. 4a, 4b, 4c illustrate three embodiments of the optical assembly of FIG. 1 and control means of FIG.

  the focus of it by varying its focal length.

  Figure 5 shows the application of the invention to a synthetic aperture radar system. Figure 6 illustrates an embodiment of the filter

Figure 2.

  Figure 7 shows the input signal of the electronic filter

Figures 2 and 6.

  FIGS. 8a, 8b, 8c, 8d and 8e show the output signal of the electronic filter of FIGS. 2 and 6 for different successive values of the order w of the filter, w decreasing, while

  the resolution increases, from figure 8a to figure 8e.

  FIG. 9 illustrates an embodiment of the Gaussian differentiation unit by spatio-temporal smoothing of

  Figures 1 and 2, with the output signals CO and DP thereof.

  FIG. 10 represents, by a curve, the histogram of the absolute values of the differences, plotted on the abscissa, between the image just before the increase of the spatial resolution and the progressively increasing spatial resolution image, with the indication the Li limit of the conserved useful portion of the histogram. FIGS. 11 and 12 respectively illustrate a still life and a human face as a perceived object, with a curve representing for each the variation of the absolute values of the above-mentioned differences (plotted on the abscissa in FIG.

  ) according to the order parameter w of the filter of Figure 6.

  FIG. 13 shows the assembly arrangement, one above the other, of the partial figures 13a and 13b to form an overall figure, hereinafter designated FIG. 13a-13b, illustrating FIG.

  the evolution, during treatment, of the image of a human face.

  FIG. 14 represents a table of the accumulation of the calculated values of CO, with DP = 1, for successive phases t of the treatment comprising a period of increase of the spatial resolution according to five stages for the order w of the filter, indicated in parentheses . Figure 15 illustrates, by curves a, b, c, d, e, the cumulative

  values of the table of Figure 14 for the different phases.

  FIG. 16 illustrates the period T3-t5 (FIG. 3) at r max and FIG.

  w = 0, with its successive spatial components a, b, c, d, e.

  FIGS. 17 and 18 show two embodiments of a device for the perception or knowledge of an object, more elaborate than those of FIGS. 1 and 2, also allowing the recognition of the object and the positioning thereof, as having an additional set consisting of a number of histogram forming units STN, Fig. 17 illustrating an elaboration of the embodiment of Fig. 1 with optical assembly for increasing the spatial resolution. in step, while FIG. 18 illustrates an elaboration of the embodiment of FIG. 2 with an electronic filter for a

such an increase.

  Figs. 19 and 20 respectively show a schematic illustration and a detailed illustration of a one-dimensional or unilinear STN (1) histogram determination unit of said additional set of

Figures 17 and 18.

  Figures 21a and 21b show two histograms

  one-dimensional, respectively at one and two peaks or maximum values, determined by the STN unit (1) of FIGS. 19 and 20.

  Figure 22 schematically illustrates a unit

  two-dimensional or bilinear histogram formation.

  FIG. 23 shows the assembly arrangement, next to each other along Z-Z ', of partial figures 23a and 23b to form an overall figure, hereinafter designated FIG. 23a-23b, which illustrates in FIG. detail such a two-dimensional or bilinear STN unit for determining histograms, denoted STN (2), of said set

  of Figures 17 and 18.

  Fig. 23c illustrates a partial variant of the CH classification subunit of the STN block (2) of Fig. 23a-23b.

  Figure 24 shows a two-dimensional histogram at

  two peaks determined by the STN unit (2) of Figures 22 and 23a-23b.

  Fig. 25 is a schematic view showing the essential elements of a self-adaptive STN module (2) with anticipation. FIGS. 26 and 27 represent two sets able successively to select classes in the case of several classes corresponding to several peaks according to FIGS. 21b and 24, the assembly of FIG. 27 constituting an improvement thereof according to FIG. to allow faster selection. Figures 28 and 28a illustrate the flowchart of the phase

  display of results in the context of the successive determination of several classes and its application in an integrated API.

  FIGS. 29a and 29b show, side by side, sets for carrying out the calculation phase, respectively without sorting

  classes and with sorting classes, in case of several classes.

  Figure 30 illustrates, in more detail, the memory 20 of Figure 29b.

  Figures 31 and 32 show, in a more

  two portions of the assembly of Figure 30.

  Figure 33 illustrates an arrangement implementing

  STN modules with multiple class extraction.

  FIG. 34 illustrates the mounting of an STN unit (2) and two planar orientation units pa, p13 in this STN unit (2) to determine a mean or barycenter position (BarZ1).

debited in a bus.

  FIG. 34a illustrates the delimitation of an object OB by a zone defined from two axes of orientation pa2 and pcc3 of the plane

  and determining the BarZ0 centroid of this area.

  FIG. 35 schematically represents bands between terminal A and terminal B, of different orientations, originating from two-dimensional units STN (2) of the set of FIG. 34 (discharging in a common bus) and passing through the BarZo center of gravity.

determined according to Figure 34a.

  Figure 36 illustrates a classification provision

  improved assembly 101 of Figure 23a-23b.

  FIGS. 37 and 38 represent the zones delimited respectively by the exit 101s of said set 101 and by the

  assembly of FIG. 36 replacing this assembly 101.

  FIGS. 39a, 39b, 39c show three successive phases of the determination of the polar coordinates p3 and c.3 defining the relative position of BarZ1, determined according to FIG. 36, with respect to the position of BarZ0 delivered by the STN unit ( 2) of

  Figure 34 and determined according to Figures 34a and 35.

  Figure 40 illustrates the father -> son relationship between BarZ0 in

  upstream and BarZ1 downstream by implementing the polar coordinates p3 and O, 3 determined according to Figures 39a, 39b, 39c.

  FIG. 41 represents an assembly constituted by an STN module (1) and an STN module (2), such an assembly being used in the

  montages of figures 42 and following.

  FIG. 42 illustrates an arrangement for determining the polar coordinates using the method illustrated in FIG.

Figures 39a, 39b and 39c.

  Figures 43a, 43b, 43c illustrate the operations

  Figs. 42 and 42 show the position of BarZ1 (the "son") relative to BarZo (the "father") in Cartesian coordinates.

  All of FIGS. 44a to 44e, 45 and 46 relate to the case of zones Z1 to Z30 that are all nested in each other; in particular: FIGS. 44a, 44b, 44c, 44d and 44e show the successive phases of the determination of the successive barycenters from BarZo to BarZ3o, when these correspond to zones Zo to Z30 nested one inside the other (as illustrated in figure 45) and thus to a dynamic tree without connections (the one

of Figure 46).

  FIG. 45 illustrates the relative positions of the BarZ0 to Bar30 centers of the zones Z0 to Z30 determined by the assembly of the

figure 42.

  FIG. 46 illustrates the non-branching shaft of the BarZ0 to BarZ30 centro-centers of FIG. 45. The set of FIGS. 47a to 47d, 48 and 49 relates to the case of zones Z10 to Z30 without any interlocking, these zones Z10 to Z30. all being included in zone Zo, in particular: FIGS. 47a, 47b, 47c and 47d represent the successive phases 1o of the determination of the successive centers of BarZ0 to BarZ30, these corresponding to zones Zo10 to Z30 without any interlocking (as illustrated figure 48) and therefore to a tree

  dynamic with connections (the one shown in Figure 49).

  FIG. 48 illustrates the relative positions of BarZo to BarZ3o centro-centers of surfaces Zo to Z30 determined by the assembly.

Figures 47a to 47d.

  Figure 49 illustrates the BarZ0 to Z30 tree of Figure 48.

  FIG. 50 shows the successive determination of zones Z0 to Z21 and Z22 in the intermediate case of nested and non-nested zones and with the corresponding tree with connection from BarZ0 and parent-son relationship between BarZ12 and Z21 on the

Figure 50a.

  FIG. 51 shows a portion of said additional set of FIGS. 17 and 18 consisting of one-dimensional units according to FIGS. 19 and 20 dealing with a parameter representative of the observed object and two-dimensional units according to FIGS. 22 and 23a-23b dealing with the parameters. x and y of coordinates. FIG. 52 shows another portion of said additional assembly of FIGS. 17 and 18 constituted by unibi- and three-dimensional units relating respectively to the value of CO, to the y and x row and column coordinates and to the L, T parameters. and S luminance, hue and saturation determining a color, the assembly of this figure 52 to retain the fugitive information determined by the assembly of Figure 51. Figure 53 shows the storage unit objects

  perceived, with translation invariance only.

  Figures 54 and 55d illustrate a unit of determination of

  invariance in size or dimension.

  FIGS. 55a, 55b and 55c correspond to FIGS. 43a, 43b and 43c respectively, but with, in coordinates, the modulus p and the angle oa, previously determined according to FIGS. 43a to 43c, and the logarithm of the distance of LD entry

determined according to FIG. 54.

  Figure 56 illustrates the result obtained by the treatment of

Figures 54, 55, 55b and 55c.

  FIG. 57 represents an improvement of the size invariance determination unit of FIG. 54, with looping of the logarithm of the distance, which allows a continuation of operation in the event of absence of a signal.

  representative of the distance.

  FIG. 58 illustrates a refinement of FIG. 47d, in addition to which FIG.

  the invariance in size according to FIG.

  Figure 59 illustrates a unit for determining invariance

in rotation.

  FIGS. 60a and 60b are partial views of the set 25 of the object recognition device with translation, size and rotation invariance, comprising units of

Figures 58, 53 and 59.

  Figure 61 illustrates the angular displacement of the

  vision to move from an object observed to another object to observe.

  Figure 62, finally, represents a scene with three objects

  identified from the point of view of their nature and position.

  We will now describe, by way of non-limiting examples,

  embodiments of a device according to the invention and assemblies and constituent units thereof, allowing the implementation of the method according to the invention.

  Although the invention is described in its preferred application implementing a video type signal, it is not limited to this type of signal; it also applies, in particular, to a

  signal issued by a synthetic aperture radar.

  A first embodiment according to the invention of a device for perceiving an object in its environment, constituting the first set of a knowledge acquisition and active visual recognition system, illustrated in FIG. first a video sensor 2, of 0o type CCD or CMOS (in particular of the type "retinal sensor"> with a large concentration of pixels in the center and decreasing concentration of pixels away from the center), forming part a camera, a camcorder or a webcam, observing an object OB located substantially in the plane 6 and responsive to a digital signal 7 having undergone a resolution reduction

  during certain periods (as specified below) and constituted by a succession of sequences (images or frames of the video signal) representative of successive views of the object in its environment and therefore falling within the time domain, each of said sequences being constituted by a succession of sub-sequences (lines of the video signal) representative of a succession of pixel locations (constituting sequenced subunits) of the video signal and therefore falling within the spatial domain, the set of locations (pixels) arranged in lines and 25 columns forming a matrix, for example rectangular.

  The video image sensor 2 is associated with an objective 5, the focal length of which can be varied, so that the plane 6 of the observed object OB can be varied, under the control of a control unit 1, three modes embodiment of the assembly 30 of the unit 1 and the objective 5 being illustrated in Figures 4a, 4b and 4c described below. This variation of the focal length of the objective 5 is controlled by the unit 1 so as to vary the spatial resolution, in the video signal 7, of the OB object to be perceived, which is in the plane 6, advantageously according to the curve of FIG. 3, in which the resolution r is carried on the ordinate and the time t on the abscissa, namely, successively after a preliminary period TO, a first period T1 of reduction, preferably abrupt, of the resolution of its maximum level, or optimum basic value, r max (of the period TO) to a minimum level, or reduced value, r min, a second period T2 of constant resolution to the reduced value r min comprising a single phase t0, a third period T3 of increasing the resolution, in stair steps, from its reduced value r min to its optimal base value r max, the increase of the resolution 1o in the third period T3 comprising several phases t1, t2 , t3, t 4, t5 of equal duration which constitute steps in the increase of the resolution, performed in a Gaussian or substantially Gaussian manner (the duration of each of these phases t1 to t5 being advantageously equal to that of a frame, but which can be equal to that of another integer number of frames), and finally a fourth period T4 completion of the process at the maximum resolution r max, also comprising several phases t6, t7, t8,

  t9, t10 durations equal to those of phases tl to t5.

  A unit 9 extracts from the video signal 7 after Gaussian transformation (FIG. 1), on the one hand, the Cartesian coordinates x

  and y of the rectangular matrix of pixels (or the polar coordinates for a circular matrix of pixels defined by a retinal sensor) and, secondly, the conventional Sync synchronization signals of a video signal (frame and line), used as explained below.

  In addition, if the video signal 7 is in color, the unit 9 extracts from it the luminance L which is applied to the unit 3. If the video signal is in black and white, the signal 7 is constituted by the signal L. The unit 3 realizes a Gaussian differentiation between two successive sequences of the video signal 7, a mode of

  preferred embodiment of this unit 3, which advantageously performs a spatio-temporal smoothing, being illustrated in Figure 9 and described below with reference thereto.

  This unit 3 finally delivers two digital signals DP exceeding and quantizing the CO overflow, explained hereinafter, the successive values of the signal CO, which is an adjustable time constant, a function of the Gaussian difference, being therefore a signal representative of the object perceived; the signals DP and CO can be viewed on an M5 monitor (that of a television or a computer for example) and / or processed in a processing unit M ', for example (but not exclusively) of the type represented in FIG. 8 in Figures 17 and 18, explained in Figures 19 and following and discussed with reference thereto; as for the DP signal, which is binary, it can, when it has one of its two values 0 or 1, perform a signal inhibition

CO, as explained below.

  In the embodiment of FIG. 2, there is the plane 6 in which the observed OB object, the video sensor 2, the Cartesian x and y coordinate extraction unit 9 (or polar coordinates) are substantially disposed. ) pixels and synchronization signal Sync, namely, for each frame, the synchronization signal of this frame at the beginning thereof, and the line synchronization signals (if the signal 7 is a color video signal, The unit 9 also extracts the luminance component L of the signal 7 ') and the spatio-temporal processing unit 3 which outputs the signals CO and DP, the output signal CO being able to be displayed on a monitor M and / or treated in the processing unit M ', monitor and assembly similar to those of the figure

  1, in the absence of inhibition by the DP signal.

  On the other hand, while the video image sensor of FIG. 2 is provided with a lens 5 with variable focus, that is to say with a variable focal length, controlled by the control unit 1 able to to vary the focus and thus the resolution, the sensor 2 of FIG. 1 is provided with a 5 'lens with a focal distance 30 kept fixed during the duration of a visual knowledge and recognition operation, the The focus is initially made on the object OB which is substantially in the plane 6. The sensor 2 thus delivers a constant resolution signal 7 ', the variation of the resolution being carried out after the extraction of the x and y coordinates and also of the the luminance L '(if the

  video signal is in color) in the unit 9 which operates on the signal 7 '(in black and white) or the luminance L' by an electronic filter 4 controlled by the control unit 1 'which imposes a variation of the resolution advantageously according to FIG. 3, by the application of a parameter w constituting the order of the filter. The filter 4 delivers a signal 7 or L, analogous to the signal 7 or L of FIG. 1, which is processed like this in a spatiotemporal processing unit 3 similar to that of FIG.

  In fact, the Gaussian or quasi-Gaussian increase of the resolution from r min to r max is carried out, in the embodiment of FIG. 1 by refocusing the objective 5 during the period T3 of FIG. defocusing during the period T1 of decreasing the resolution from r max to r min and constancy of the resolution to the value r min during the period T2), whereas, in the embodiment of FIG. 2, this increase is carried out by the electronic filter 4 which electronically performs the equivalent of a refocusing during the period T3 (after the equivalent of a defocus in the period

  T1 and maintenance of the resolution in the period T2).

  We will now indicate, with reference to FIGS. 4a, 4b, 4c, three embodiments of the variation of the optical resolution, by variation of the focus of the objective 5 of FIG. 1. In FIG. found the plane 6, the OB object, the video image sensor 2, the variable focal length objective, noted 5a, the control unit, noted la, the adjustment of the focal length of the lens of Figure 1, this unit acting on the position of at least one of the lenses (or single lens) of the lens 5a; the change of position of the lens by two-way displacement, which is schematically illustrated by the double arrow f, modifies the focus of the lens relative to the plane 6, with a substantially Gaussian increase in steps of the resolution during the period T3 of increase thereof from r min to r max, after the abrupt decrease in resolution to r min during the period T1 and the maintenance at r min of the resolution

during the period T2.

  FIG. 4b shows the plane 6, the OB object, the video image sensor 2, the variable focal length objective, denoted 5b, the control unit, denoted 1 b, of distance adjustment. focal point of the objective of Figure 1, this unit lb acting on the focal length of the lens 5b by means of an electric control current whose variable voltage U can adjust the focal length of the lens 5b to vary the resolution according to FIG. 3. The set lb-5b can be, for example, of the type described in the French patent application having the publication number.

  2,769,375 (filed October 8, 1997 under number 97 12781).

  In place of the modifiable lens making the variation of the resolution of the objective 5b of FIG. 4b, it is possible, as illustrated in FIG. 4c, to provide the combination of a lens 5 'of constant focal length and a lens. hollow lens 5c with variable focal length disposed in front of the 5 'lens, this lens 5c being controlled by a control unit 1c which regulates the arrival of a transparent fluid inside this lens 5a between two sheets 51 and 52 which are substantially parallel to the rest and which bulge under the action of the arrival of said fluid by changing the focal distance of this

lens 5c.

  The three means of variation of the focal length, therefore of the resolution, of FIGS. 4a, 4b and 4c result, because of the laws of optics applied to the objective 5 (of the type 5a or 5b) or to the the

  5c to Gaussian, or at least quasi-Gaussian, the resolution of the optical system and therefore results in a signal L or a signal 7 which constitutes the Gaussian transform of the video signal, especially during the period T3 .

  FIG. 5 illustrates the application of the invention to the processing of the output signal 7R (analogous to the output signal 7 or L of FIGS. 1, 4a, 4b, 4c) of a radar transceiver 2 '. with a synthetic aperture emitting a lobe, referenced Lob, with variable aperture, whose finest lobe corresponds to the maximum resolution r max, while the widest lobe corresponds to the reduced resolution r min (as indicated in FIG. 3 ). A control unit 1d controls a unit 5d of synthetic opening of the emission lobe to adjust, on the one hand, the width of the emitted lobe and, on the other hand, its matrix scanning, normally according to a corresponding rectangular matrix. to a matrix of video pixels of this type. As a result, the signal 7R emitted by the transceiver 2 'examining the object OB in the plane 6 is very similar to the signal 7, or rather L, of FIGS. 4a, 4b and 4c and can therefore be processed so similar to this signal 7, or rather L, in the unit 3 of Figure 1; it will be noted, in fact, that this signal 7R of FIG. 5 has only one digital component, unlike the signal 7 of FIGS. 4a, 4b, 4c which, in the case of a color video sensor 2, comprises three color components, this signal 7R is rather analogous to the luminance signal L of FIG.

  also a single component signal. As a result, the unit 9 '(corresponding to the unit 9 of FIGS. 1 and 2) produces only the extraction of the synchronization signals Sync from the radar signal 7R (in addition to the x and y coordinate signals not indicated on Figure 5).

  It is also known that the signal 7R, because of its obtaining as an output signal of a synthetic aperture radar, has the form of a Gaussian or possibly quasi-Gaussian transform. Similarly, a Gaussian or quasi-Gaussian processing is performed by electronic filtering, in the filter 4, in the case of FIG. 2 (as specified hereinafter), as a Gaussian or substantially Gaussian modification of the setting is made.

  optical point in the case of Figure 1.

  We will limit ourselves now with reference to FIGS. 6, 7 and 8a to 8e to the embodiment of FIG. 2 with Gaussian filter 4.

  or quasi-Gaussian to realize the variation of resolution.

  The electronic filtering unit 4 constituting the filter of FIG. 2, which in particular varies the input signal 7 'during the period T3 of incrementally increasing resolution (FIG. 3), is constituted by the succession of two successive filtering units for the x and y coordinates, each of these two units being either of the Gaussian type or of the substantially Gaussian type and, in this case, be constituted by a Canny filter, discussed for example in an article of Messrs. Didier DEMIGNY and Tawfik KAMLEH entitled "A discrete expression of Canny's criteria for step-by-step detection performance evaluations" in IEEE Pattern Analysis and Machine Intelligence, Volume 19, No. 11, pp. 1199-1211, November 1997, or even better by a filter 10 of the PAOG type (Polynomial Approximation of Gaussian) designed by Didier DEMIGNY, Julien PONS, Nassima BOUDOUANI and Lounis KESSAL; In both cases, these are response filters.

impulse finite.

  The electronic filtering unit 4 is designed to perform a substantially Gaussian or even Gaussian filtering.

  following the column x coordinates and y row of the pixel array; advantageously, this filtering is carried out in two steps, namely preferably first filtering according to y in a unit 20, then filtering according to x in a unit 21, as illustrated in FIG. 2 inside the filter 4 .

  The filter unit 20 is advantageously of the latter type mentioned above and is illustrated in FIG. 6. It comprises, as constituent blocks, one-storey Re registers, represented by rectangles, with long vertical sides, imposing a delay. unit, that is to say a sub-sequence or a row of pixels, two registers A and B, represented by quasi-square rectangles detailed in ab, imposing a delay of w lines (w representing the order of the filter, which is progressively modified as explained below), subtracters ns, mu multipliers and adders ad, respectively denoted by the classical symbols -, x, +, the input signals ao, a1 and a2 of the three multipliers mu of the subunit 20a being calculated, from w, by the formulas a0 = w, a1 = w + 3 and a2 = 2w + 3, while the input signal Cp of the multiplier of the subunit 20c is given by the formula: Cp = 5 / {2w (w + 1) (w + 2) (w + 3) (2w + 3)} The signals ao, a1, a2 and Cp (normalization coefficient), which are only a function of the order of the filter w, are calculated by the four above-mentioned formulas in the registers A and / or B 5 (ab) imposing a delay of w subsequences or lines of pixels as a function of the signal w (variable as explained below) received in

  input, as shown in ab in Figure 6.

  All of said building blocks, namely the registers and

  the arithmetic operators of the filter unit 20 form three subunits, namely a non-recursive subunit 20a, a recursive subunit 20b and a smoothing subunit 20c; in subunit 20b, each register pair Re + adder ad preceding it, with loopback, constitutes an integrator In, while, in subunit 20c, the register pair Re + adder ad following it, with looping, constitutes a smoothing unit Ls.

  The filter 20 of FIG. 6 finally delivers the quasi-Gaussian transform at y of the input signal L '(or 7'), this signal L '(or

  7 ') being the one at the inlet of the total filter 4 (Figure 2).

  The unit 21 of this filter 4, which has produced an almost Gaussian processing in x, is identical to the unit 20 illustrated in FIG. 6 and which has just been described, apart from the fact that the input, instead of be constituted by the signal L '(or 7'), is constituted by the output sy in y of the unit 20, while the output of the unit 21 is constituted by this transform in y (debited by the unit 20) having in turn undergone a quasi-Gaussian transformation in x; in addition, the unit shift of the registers is this time not a substequence or row of pixels, but a pixel position for the registers Re, while the offset is w pixel positions for the registers. Registers A and B. The output of the unit 21, and thus the output of the total filter 4 of FIG. 2, is a signal L (or 7) analogous to the signal L.

(or 7) of Figure 1.

  The impulse response of the filter set 4 (the succession of the units 20 and 21) is: h (w, k) = Cp. (W + 2-Ik1) (w + 1-lk1) {- 3k2 + (2w +3) lkl + w (w + 3)}, with Cp 35 given by a formula above, along the two axes y and x of the image, while w, which constitutes the order of the filter 4, is given by the a = 0.3217w + 0.481 (w being substantially equal to 3a for relatively high values of a); it is recalled that a is the standard deviation of the Gaussian and that the number of coefficients of the impulse response of the filter, related to the resolution, is equal to 2w + 1, the value chosen for Cp having as its object to make equal to

1 the sum of said coefficients.

  FIG. 7 illustrates the signal 7 ', namely the output signal of the video sensor 2 of FIG. 2, possibly reduced to the 10 luminance signal L' in the case of color video, for a particular line of a given frame of the video signal; in this figure, the x-coordinate representing the successive pixels of the line is plotted on the abscissa, with the column number being indicated in the rectangular matrix constituted by all the pixels of a frame arranged in columns and in lines, while that ordinate, the level ni of the signal 7 ', in particular the luminance L', which, it is recalled, is at the maximum resolution r max (FIG. 3) or base resolution, before having undergone a transformation

  Gaussian in the filtering set 4 of FIGS. 2 and 6.

  On the contrary, in FIGS. 8a, 8b, 8c, 8d and 8e, with the same abscissae x and the same ordinates as in FIG. 7, the signal 7 or L after filtering is represented in the filtering unit 4. for different values of the order of the filter, namely respectively w = 20, w = 16, w = 12, w = 8, w = 4, that is to say 25 in order of decreasing filter in steps, so by increasing resolution (this can be seen in Figure 3). We note that we chose a law of arithmetic decay for the variation of the order of the filter, from one figure to another of the set

Figures 8a to 8e.

  Comparing FIG. 7 with the various FIGS. 8a to 8e, it can be seen that, before filtering (period TO), the signal 7 'to r max, with C = 0 (FIG. 7), comprises all the variations (even the short ones) of the level nor the signal, in particular the brightness, representative of the object seen by the video sensor 2, whereas, when the resolution is abruptly reduced from its maximum value r max to its minimum value r min reached at the end of the period T1, the signal 7a of the period T2 when w = 20 (Figure 8a) is highly smoothed and therefore only certain characteristic areas are vaguely reproduced; as resolution 5 rises in steps from its r min value to its r max value, the value of the decreasing filter order (FIGS. 8b to 8e) to go from w = 20 to w = 16, and then at w = 12 and w = 8 and finally at w = 4, more and more fine details appear in the curves 7b to 7e,

  FIG. 8e being relatively close to FIG.

  Thus, FIG. 7 represents the signal 7 '(in particular L') at the maximum resolution, before filtering, whereas FIGS. 8a, 8b, 8c, 8d and 8e represent the signal 7 (in particular L) resulting from the filtering of the signal. signal 7 '(in particular L') by the filtering unit 4 of FIGS. 2 and 6. It can therefore be seen that the output signal which, before modification of the resolution, was actually representative of the OB object with its details , undergoes a sharp reduction of the resolution (equivalent to a defocusing as envisaged in the case of FIG. 1), then little by little (FIGS. 8a to 8e) sees its resolution increase stepwise (as illustrated in FIG. 3), To arrive at the curve 7e of FIG. 8e, which reproduces, with a certain smoothness, the initial curve 7 'of FIG. 7; Moreover, if we compare the set of FIGS. 7 and 8a-8e and FIG. 3, we see that FIG. 7 corresponds to the period TO of FIG. 3, whereas FIGS. 8a to 8e correspond to the phases (FIG. or 25 steps) t0 to t4 of the periods T2 and T3 of FIG. 3, w decreasing progressively until reaching the value zero corresponding to r max, that is to say to the resolution before filtering or defocusing,

at the t5 phase of the T3 period.

  Finally, the output, noted by simplification L or 7, of the filter assembly 4 is constituted by the succession of the signals 7a, 7b, 7c, 7d and 7e (of FIGS. 8a, 8b, 8c, 8d, 8e respectively ) during the phases t0 (period T2), t1, t2, t3, t4

(of period T3) of FIG.

  The signal L (or 7), or 7R, i.e. the Gaussian or quasi-Gaussian transform of either Figure 1 or Figure 2, or Figure 5, is processed in a unit 3 of FIG. This unit 3 is of the type described and illustrated for the first time in the international publication WO98 / 05002 (or rather its priority document) having the same inventor as the present application, particularly in the context of FIG. passage on page 14, lines 24 to 21, line 19

  and to which one can possibly refer.

  The unit 3 of FIG. 9 firstly comprises two subunits 10 and 11, the subunit 10 being a memory, while the subunit 11 (corresponding to the subunit 15 of FIG. last mentioned publication) performs a spatio-temporal smoothing treatment. It can be seen that the parameters LO and CO circulate in loop by being reinjected, from the outputs of the subunit 11, to the inputs thereof, after having undergone, in the memory 10, a delay equal to one sequence, namely an image or a frame in the case of a video signal. As a result, thanks to the memory 10, which stores the values of LO and CO during a sequence, the unit 11 compares the values of LO and CO earlier than a sequence with their current values, for each pixel the 20 letters t, tI, x and y respectively representing the instant t, the instant t-1 (t-1 corresponds to the instant preceding the time t of the duration of a sequence), the abscissa x and the y-ordinate of the pixels, t corresponding to the current sequence, while t-1

* corresponds to the previous just sequence.

  The recursive calculation function in unit 3 of FIG. 7 corresponds to LOtxy = LOt-1, x, y + (Pixt, x, y-LOt-1, x, y) / 2cOtxy with the following test: - if IPixt, x, y-LOt-1, x, y I> threshold, then DP = 1 and, for 0 <CO <p, COt, x, y = COt-lx, y -1 - otherwise,

  then DP = 0 and, for 0 <CO <p, COt, x, y = COt-1, x, y + 1-.

  LO (t-1) represents the substantially anterior Gaussian image, smoothed in the subunit 11 and delayed by a sequence in the memory 10, while Pixt is the current Gaussian image represented by the input signal L ( or 7), or 7R, LOt thus constituting a difference of Gaussian (called DOG), spatio-temporal; LO represents the successive smoothing values, while Pix represents the pixels of the L or 7 signal, or the values of the radar signal 7R, at different positions in the pixel distribution matrix or point radar signals; finally, CO and DP are the output signals of the smoothing set

spatio-temporal 3.

  Whereas, in the prior international publication, the input signal consisted of a conventional video signal delivered by a video sensor, here we have the signal L (or 7) of FIGS. 1 and 2, or the signal 7R of FIG. 5, namely the input signal Pix of the unit 3 of FIG. 9, more particularly on one of the inputs 5is of the subunit 11 of this unit 3 which calculates a Gaussian difference. The CO and DP output signals of the unit 3 result from the comparison of a characteristic parameter of the video signal 7, for example the luminance L of the pixels thereof, for the same pixel location, between the value for the frame just prior and the value for the current frame of the video signal, the adaptive time constant CO tending to reduce this difference as shown by the group of formulas last quoted. It is the same in the case of an electromagnetic signal

radar type.

  The overflow signal DP is a binary signal, the two values of which represent respectively the exceeding and the not exceeding of a sensitivity threshold (referenced threshold at 12a and which is greater than the variations of the input signal L (or 7) or 7R due to the background noise) by the difference in absolute value, for each same pixel location, between two consecutive sequences (images or frames), for example at the instants t = 0 and t = 1 for the images lca and 113 respectively , while the digital signal CO, which has a small number of bits (in the monomer 2co with CO = n, n being this number of bits), represents an adaptive time constant reinjected into the process in order to reduce the number of bits.

  variation between two consecutive sequences, for the same pixel location, of the processed signal; the values of CO, when DP has the value representative of an exceeding of the threshold, are themselves representative, as a function of the successive sequences s of the difference of Gaussian and therefore of the increase of the resolution, as illustrated by the picture ly.

  Unit 3 is associated with a subunit 14 constituted by an STN module (1), of the type described hereinafter with reference to FIG. 20, which forms the histogram of the difference in absolute value 10 (explained above). with reference to Figs. 19, 20, 22 and 23 and denoted by Dif) between Pix t, x, y and L0t-lxy, calculated in the smoothing subunit 11, which will be used as explained hereinafter with

  reference to Figures 10, 1 1 and 12.

  FIG. 9 illustrates, on the one hand, the images loc and 1f3 more or less blurred, of the same face, conveyed by different input signals of the subunit 11 of spatio-temporal smoothing, at times t and t1, namely t = 1 and t = 0, for orders w of the filter of 16 and 20 respectively, in the case of the device according to FIG. 2 with electronic filter, the resolution r increasing (as illustrated in FIG. 3) during the period T3 when the order w decreases from 20 to 16 (from the lac image to the image 11); when DP = 1 (value chosen to represent an exceeding of the threshold), the higher or lower value of CO, namely n in the monomer 2co = 2n, (n being a low numerical value), then makes it possible to reconstitute an image more or less vague as explained below. The image ly of output of the subunit 11 of spatiotemporal smoothing, at time t = 1, with w = 16, displayable on the monitor M of FIGS. I and 2, is deduced from the value, at this moment, of CO, or rather n, which decreases as a function of I Pixt, x, y-LOt-11, x, y 1, when COt, xy = COt-1, xy-1 and DP = 1, which corresponds to has a

  exceeding the sensitivity threshold by said absolute value.

  It should be noted that, in the case of the device of FIG. 1, with variation of the resolution by modification of the focusing, and that of FIG. 5 with variation of the resolution by modification of the radar lobe, is the increasing value of the step resolution, refocusing or decreasing the aperture of the radar lobe respectively, which is substituted for the decreasing value

  in steps of the order w of the filter 4 of FIG.

  At the spatio-temporal smoothing unit 3 is associated the control unit 1 (FIG. 1) or FIG. 1 (FIG. 2), detailed in FIG. 9, which first comprises an STN module (1). 14, of one-dimensional Reg-register type, for example that described with reference to FIG. 20, processing, as input data (DATA A), the signal Dif = I Pixt, x, y-L0t-1, x, yl , this absolute value of the difference between two successive images (Img) at times t and t-1

  (the unit value 1 corresponding to the duration of a sequence of the signal L, 7 or 7R) being also represented by Imgt-lmgt.il (according to Figure 10) the difference between two successive images (lmg) at times t and tl (The unit value 1 corresponding to the duration of a sequence of the signal L, 7 or 7R.

  Dif signal processing in module 14 provides

  the histogram of this signal, as shown in Figure 10.

  The control unit 1 or 1 'also comprises a control box 15 for the STN module (1) 14, whose start is controlled by the initialization signal Dep and which, under the control of the synchronization signal Sync common to the entire device of FIG. 1 or 2, outputs the period selection signals representative of the successive periods of FIG.

  know the periods T1, T2, T3, T4, for the module STN (1) 14.

  Once the control box 15 has been triggered by the signal Dep and it delivers the successive signals of the phases t1, t2, t3, t4, t5, t6, t7, etc. which determine the successive values of w in the embodiment of FIGS. 2 and 6 (t1 to t5), or the successive resolutions of the objective 5 of FIG. 1 (in particular the objective 5a, 5b, 5c of FIGS. 4a, 4b, 4c respectively) , or the synthetic aperture unit 5d of a radar determining the successive apertures of the Lob radar lobe (FIG. 5), these period signals T1, T2, T3, T4 with their phase signals t1, t2, etc. ... control the operation of the STN module (1) 14 which determines a histogram of the parameter Dif

  received from the smoothing subunit 11.

  FIG. 10 shows the histogram (formed in the STN module (1) 14 of FIG. 9) of the difference (in absolute value) Dif = 1 Pixw0, x, y-Pixwix.yl or simply Dif = 1 Imgwo-lmgwi l, Img being an abbreviated notation of the word image, while wO and wi are the initial and current values of w. In this FIG. 10, the number of pixels corresponding to this difference in absolute value plotted on the abscissa is represented by N on the ordinate. From the curve of FIG. 10, a limit value Li of these differences Dif which makes it possible to group for example 90% (it is advantageous to choose in any case a value equal to at least 75%) of the points of this criterion, is determined. difference, choosing for example w = 0 and w = 20, that is to say the two values of w 15 corresponding to r max and r min. The difference corresponding to the

  Li limit represents the error between the same points of two images, the original first, unfiltered rank 0, delayed by the subunit 10, representative of the sequence TO, the second, L (7) filtered rank corresponding to Li in Figure 10. This makes it possible to finish working on the most stable elements of the image.

  Since the limit Li has been determined from the difference distribution histogram Dif (in FIG. 10), this limit is considered to be the maximum variation caused by the filtering at w max with a number of intervals of I such that Li The threshold 25 referred to here is the sensitivity threshold which appears in 12a in the unit 11 of FIG. 9, this threshold being greater than the background noise in order to distinguish two successive stages. The value of p is forced to I (number of intervals) ie in the particular case illustrated. By way of example (ie the illustrated example), if we found Li = 27 (90% of the differences) and if the threshold value is 5, then we take I = 5, because we have 27/5 > 5 and p = I = 5. In the case of a non-integer result of the division of Li / threshold, round to the lower integer. Consequently, in the example chosen, there will be five stages, starting from the phase t0 (period T2), namely the five phase bearings t0, t1, t2, t3, t4, of FIG. curves 7a, 7b, 7c, 7d, 7e of FIGS. 8a, 8b, 8c, 8d, 8e respectively. The maximum value for w, in this particular case, being 20 (phase t0), we deduce the difference w between two consecutive levels, Aw = 20/5 = 4. Thus with a maximum w equal to 20, we use the following parameters for the successive frames: to w = 20 n = 5 n = p minimum resolution r min t1 w = 16 n = 4 2co = 2n t2 w = 12 n = 3 t3 w = 8 n = 2 t4 w = 4 n = 1 t5 w = 0 n = 0 Returning to FIG. 9, it can be seen that in unit 3 of differentiation 15 of Gaussian is associated a subunit 16 for calculating the

  number of steps, default integer estimated value of Li / threshold, Li from the STN module (1) 14 and being deduced from the histogram of FIG. 10 determined by the module STN (1), while the threshold (of sensitivity ) originates from the location 12a of the subunit 11 20 O it has been previously registered.

  The value of p determined by the subunit 16 acts on the

  control 17 of the unit 4 (Figure 2) to vary the order w of the filter 4, so the spatial resolution determined by this filter (or the spatial resolution of the objective 5 of Figure 1, or the unit 5d controlling the opening of the radar lobe in the case of Figure 5).

  Finally, FIG. 9 illustrates, on the one hand, the control API of the

  subunit 14 of subunits 16 and 17.

  Practically, after the abrupt decrease in the resolution (period T1), which passes from r max to r min (FIG. 3), the input signals of the subunit 11 (FIG. 9) are constituted during the phase t 0 (period T2), by the resolution image r min conveyed by the signal 7 (or L) or 7R, which enters as Pix (video signal) or analog signal of the electromagnetic radar type and by the previous resolution image r max who comes back as an LO. The calculation of the absolute value I Pixwoxy-Pixwi, x, y, making it possible to obtain the histogram of FIG. 10, is used at the end of the sequence to determine Li, then Li / threshold and finally the number of steps, at know p

(Li / threshold> p).

  During this phase t0, it is essential to memorize the image at r min, as being the reference during the resolution increase, by forcing p = 0 via the multiplexer 13; LOt, xy = Pixtxy. LOtxy is the filtered image of rank corresponding to

  w = 20, it becomes the image la in the following phase tl.

  In the following phase t1 (at the beginning of the period T3) 1o corresponding to the first increase step of the resolution starting from r min, the value p = 5 is forced through the multiplexer 13 (in the case particular selected) as a CO input signal; the input signal Pix corresponds to said first step, while the signal LO corresponds to the preceding resolution signal r min. We then obtain output signals of the smoothing subunit constituted by DP = 1 and CO = p-1 = 4, as

indicated for Figure 9.

  Thus, the calculation of p is carried out in the subunit 16 and the value of p thus determined is written at 12b, then, firstly, forced during the phase t1 in the subunit 11 as C0t_ to COt.o, via the multiplexer 13 and, secondly, transmitted to the control sub-unit 17, which gives, in response, the order of execution of the successive levels of increase of the resolution. from r min to r max in five steps (FIG. 3), either at 25 the objective 5 to vary its focal length (FIG. 1), or at the filter 4 at y for a gradual decrease in w (FIG. 2), or at l 5d synthetic aperture unit for varying the aperture of the radar lobe

Lob (Figure 5).

  FIGS. 11 and 12 show the histograms 30 of the values of the difference Dif in relation to corresponding images which illustrate, on the one hand, a still life (FIG. 11) and, on the other hand, a human face (Figure 12). In these two figures, the horizontal axis of the abscissas of the histograms represents the values of w, thus the level of reduction of the resolution which evolves in the opposite direction, while the vertical axis of the ordinates represents the difference in absolute value, noted dif,

  resumption of abscissas in Figure 10.

  All of FIGS. 13a and 13b, arranged one above the other (as shown in FIG. 13) illustrate the progress of the processing steps. The first image at the top left of FIG. 13a, subtitled IMG, is the image of FIG. 9 corresponding to the signal 7 'at the maximum resolution r max before the reduction of the resolution; it is a complete picture well legible. As for the image at the top right of Figure 13a, 0o it corresponds to the left image with maximum defocus (or resolution r min) during the phase tO, after application of the value 20 for the order w of filter 4. This image, which is the IP image at the tO phase, or lao at the tl phase of Figure 9, eliminates a large

  part of the details of the previous IMG image.

  After the abrupt reduction during the period T1 (Figure 3) of the resolution which passes to its minimum value r min, the details reappear and we thus obtain for w = 16 the subtitled picture Iw (16) which corresponds to the Fig. 9, in phase t1, and for which n = CO = 4 (see the preceding small table), DP being equal to 1 as for the following images of Figs. 13a and 13b (in effect, when DP = 0, there is no overshoot and therefore there is inhibition of the signal CO, as symbolized on the right in FIGS.

and 2).

  We see that as w decreases, and goes from 16 to 12, then to 8, then to 4 and finally to 0 for the image Iw (0) 0, details

  more and more numerous and more and more fine appear.

  The following images Iw (0) 1, Iw (0) 2, Iw (0) 3 and Iw (0) 4 (from the bottom of Figure 13a and Figure 14) correspond to the phases t6, t7, t8 and t9, that is to say at the period T4 of FIG. 3. Since the value of CO as long as it is not zero, it is not possible to instantly correct the difference between two successive images, and we obtain, after t5, images which progressively fade from Iw (0) 1 to Iw (0) 4 during the whole period T4 during which the information progressively decreases, this period making it possible, on the other hand, to consolidate the previous treatments

in unit 3.

  The following table illustrates the evolution of the system, sequence after sequence, that is, frame after frame in the case of a video signal, during recursive smoothing. w N values of N for ti n

0 1 2 3 4 5

  to w = 20 N = p = Aw -1 LOtxy = 0 0 0 0 0 0 Pixtx, t1 w = 20 - Aw = 0 0 0 0 a 5 _ 16 t2 w = 12 0 0 0 ab 5 t3 w = 8 0 0 abc 5 t4 w = 4 0 abcd 5 t5 w = 0 abcde 5 t6 w = 0 bcde 0 5 t7 w = 0 cde 0 0 5 t8 w = 0 from 0 0 0 5 t9 w = 0 e 0 0 0 0 5 On this table a, b, c, d, e correspond to numbers of pixels for n <p with DP = 1. The value a represents a very low spatial frequency l 0 and the values b, c, d and e correspond to filtering of spatial frequencies of higher and higher. At the end of t5 (that is to say of + p), we have the maximum of information for all the levels, in the same frame,

  as shown in Figures 15 and 16, discussed below.

  Figure 14 shows additional information from the previous table, including the indication of the periods, with the initial period T1, with X for the previous values of n or CO, and the numerical values representing the cumulative pixel variations per frame for a, b, c, d, e, increasing from bottom to top left

and to the right.

  FIG. 15 illustrates, by curves a ', b', c ', d', e ', the successive values of a, b, c, d, e of FIG. 14, the time t 5 being on the abscissa and the number N equal to the sum of

  pixel variations containing CO, with DP = 1, as ordinates.

  The sum of the areas under each of the curves a ', b', c ', d', e 'gives the total information on the image; we find that the sum of the areas under the first two curves a 'and b' 10 represents almost half of the total information, of o possibility

  to stick to these two curves.

  In FIG. 16, the images corresponding to the values of a, b, c, d, e (these letters being carried at the bottom of the corresponding images) are illustrated at time t = p and it can be seen that the details are of more and more important as we go from a to b, from b to c, from c to d and finally from d to e, the image which is at the top left, subtitled t = p, corresponding to the superposition of the images a, b, c, d and e; it is similar to

  the image labeled Iw (0) 0 in Figure 13a.

  We have seen so far that, thanks to appropriate treatment

  of the signals CO and DP of FIG. 2, it is possible to have increasingly fine and increasing details of the object OB observed by a stepwise increase in the spatial resolution, according to FIG. after a sharp decrease in that one.

  The modification of the resolution is performed in the case of Figure 2 by means of a filter assembly 4 (detailed in Figure 6) controlled by a parameter w constituting the order of the filter; such a Gaussian or substantially Gaussian transformation, in the filter assembly 4, is equivalent to the optical defocusing performed in the embodiment of FIG. 1 and therefore the signals 7, essentially of luminance L, and the signals The numbers CO and DP are the same in the case of FIG. 1 and in the case of FIG. 2, and therefore substantially the same images at the output of the spatio-temporal smoothing unit 3 of FIGS. 1 and 2 are obtained. that is to say that FIGS. 8 to 16 and the explanations relating to them also apply to the case of a modification of the spatial resolution by optical defocusing, then by step refocusing (for example 5 steps), rather than by electronic filtering. It is the same when the optical sensor of FIGS. 4a to 4c is replaced by a sensor

  electromagnetic radar type according to FIG.

  In the case of a defocusing followed by a refocusing, both of optical nature according to FIGS. 1, 4a, 4b, 4c, it can be very difficult to carry out a refocusing in net steps according to FIG. 3 and, to overcome this difficulty, it is advantageous to associate an electronic shutter (not illustrated) of known type with the objective 2 in order to use only reduced portions, regularly spaced from the signal while the refocusing is substantially constant 15 to refocus

  in steps similar to that obtained by electron filtering according to FIGS. 2 and 6. It is understood that the openings of said shutter take place outside the calculation cycles of the STN modules, which will be mentioned hereinafter with reference to FIGS. and following.

  This summarizes the operations carried out in

  devices of the embodiments of Figures 1, 2, 4a, 4b, 4c, 5, 6, 9 with reference to Figures 3, 7, 8a to 8e, 10, 11 or 12, 13a13b, 14, 15 and 16.

  From the start or start signal Dep (indicated in FIG. 3 at the beginning of the period of T1 and FIG. 9 at the input of the module): In the first period T1, which lasts, for example, a sequence (picture or frame), the signal L, 7 or 7R undergoes a very rapid reduction of its resolution from r max to r min, while w passes from 0 to 20 for example (FIG. 3), and the unit 3 of spatio-temporal smoothing (FIG. or 2 and 9) has memorization of

  the unfiltered image of the TO period.

  The following period T2 is constituted by the single phase t0, of duration equal to one sequence for example, as the phases t1 to t9 of the periods T3 and T4 (FIG. 3). During this phase t0, the resolution of the signal L, 7 or 7R remains constant at the value r min, w being equal to 20 for example and the value p = 0 is forced in the subunit 11; the memory sub-unit 10 of FIG. 9 records the minimum resolution signal and outputs, among other things, a signal

  L0to which, for its different values at the locations x, y of the pixel matrix (or similar positions in the case of an electromagnetic signal of the scanning radar type) is represented by the unfiltered image of rank 0 in, corresponding to signal L, 7 or 10R (image L (7) 20 of FIG. 13a) is introduced into the unit 11.

  The unit 11 delivers a signal Dif for the STN module 14 to establish the histogram of FIG. 10 and to determine at the end of the phase t0 the value Li transmitted to the subunit 16 which deduces, from the threshold of sensitivity "threshold", the value of p, namely the integer 15 (default) resulting from the division of Li by the threshold value [abbreviated p = int (Li / threshold), the symbol int representing the integer value , by default, the quotient, as shown in Figure 9]. For example p = 5 and this value is forced into the sub-unit

11 during the next phase tl.

  At the beginning of the next T3 period, i.e. at the beginning of the tl phase, the resolution of the L, 7 or 7R signal is scaled up and begins the normal routine operation of the unit 3 of spatio-temporal smoothing with LO, CO, DP and Dif calculations from subunit 11; during this phase t1, the incoming picture in this subunit is the picture I3 (Fig. 9); this image is calculated from the incoming image of the previous phase corresponding to the minimum resolution image, namely the image la (FIG. 9). On the other hand, the outgoing image, represented by the CO (with DP = 1), is the image ly (FIG. 9) or Iw (16) (FIG. 13a) which has very few details, but the most

  representative; during the following phases t2, t3, t4 and t5, the process continues successively with w = 12, 8, 4 and 0 and the images Mw are obtained on the monitor Mw (12), Iw (8), Iw (4) and Iw (0) 0 respectively of Figure 13a. It can be seen that during the t5 phase, the resolution has regained its maximum value r max, with w = 0.

  During successive successive phases t6, t7, t8, t9 of the subsequent period T4, on the one hand, the resolution remains at its value r max, and, on the other hand, the order w of the filter 4 remains equal to 0 (FIG. 3) in the case of the embodiment of FIG. 2, or else the focal length 5 of the objective (in particular 5a, 5b, 5c) has made a refocusing on the plane 6 of the object OB, or the control plate 5d has imposed on the lobe Lob of FIG. 5 its minimum opening corresponding to a refocusing or to w = 0. The process in the unit 3 continues to proceed because CO only goes back down progressively 10 unit value, the exponent n of 2n decreasing progressively 4 to 3, then to 2, then 1 and finally 0, which gives the images, Iw (0) 1, Iw (0) 2, Iw (0) 3 and Iw (0) 4 of Figures 13a

(last row) and 13b.

  At the end of the period T4, that is to say of the phase t9, the filter 4 s15 or the objective 5 or the lobe Lob returns to its state of rest at the maximum resolution r max; a new cycle, according to Figure 3,

  can start again from a new signal Dep.

  The first simplified embodiments of FIGS. 1 and 2 of a device according to the invention having been discussed above, reference will now be made to the following figures, and first of all to FIGS. more elaborate embodiments using STN modules (in addition to the histogram calculation module STN (1) 14 of FIG. 9), downstream of the spatio-temporal smoothing unit 3. Note that Figure 17 derives from Figure 1 to which was added a unit 8, while Figure 18 derives from Figure 2 to which was added a

same unit 8.

  The difference between FIG. 17 and FIG. 1, on the one hand, and between FIG. 18 and FIG. 2, on the other hand, is constituted by the fact that the CO and DP signals originating from the aforementioned unit 3 are treated in a set 8 which carries out the formation and the classification of histograms by using modules of the type

STN, explained below.

  It should be noted that, in the case of the embodiments of FIGS. 1 and 2, the spatio-temporal smoothing unit 3, as detailed in FIG. 9, does indeed comprise an STN type unit, but this unit processes the Dif signal, to obtain the histogram of the figure

  , and not the signals CD and DP, of the subunit 11.

  It is understood that the embodiments of FIGS. 4a, 4b and 4c apply as well to the assembly of FIG. 17 as to that of FIG. 1, whereas the embodiment of FIG. both in the assembly of Figure 18 and in the

of Figure 2.

  The other parameters processed in each of the units 8 10 are, firstly, two position parameters, noted, x, y, and, secondly, three parameters representing a color, namely the luminance L, the color T and the saturation S, these five parameters having been extracted by unit 9 'similar to unit 9 (which in the embodiments described above extracted only the parameters x and y) of FIGS. 1 and 2; of course, instead of the L, T parameters

  and S, one can represent a color by other triads of parameters, for example parameters representing the fundamental colors R red, green V and blue B or, keeping the brightness L, instead of T and S, the two parameters CR, which corresponds to LR, and CB, which corresponds to LB.

  If the L, T and S parameters have been chosen, it is to take advantage of the fact that, in the event of a change of lighting, the parameters T and S are practically not modified, only the luminance L being a direct function of illumination, which means that only one parameter is modified and, as explained, it is this parameter of luminance L which is favored compared to the two others, whereas, if one used triads of parameters of the type red, green and blue or type L, LR and LB, the three parameters are changed in case of change of illumination. In fact, luminance is essentially used in a first

  processing phase in the set 8, as will be seen hereinafter, while the other two color parameters, namely T and S are used essentially to maintain "hooked" the image when it was obtained by treating only L (with of course simultaneously x and y).

  A further input parameter in the set 8 denoted Dis, equal to the distance measurement between the object OB and the sensor 2, resulting from an external unit 25 (not described), assists the perception of the object by a size invariance calculation by STN modules described later. The output of the set 8, constituted by STN type modules, delivers two groups of signals making it possible to identify the nature of the object (signal QUOI or LABEL) and the position of the object

(signal O or Z0).

  The set 8, whether in FIG. 17 or in FIG. 18, as explained below, is the same since, as explained above, the variation of the resolution, be it made by optical means (FIGS. 1 and 17) or by electronic filtering (FIGS. 2 and 18), leads in both cases to the same type of increase in the resolution, namely Gaussian or substantially Gaussian, during the period T3, unit 3 of FIGS. 17 and 18 finally receiving the same type of signal 7, essentially of luminance L, which is spatiotemporally processed in this unit, and delivering CO and DP signals of the same kind in the case of FIGS. 1 and in the case of FIGS. 18 and 2: the signal CO (which is an adaptive time constant appearing in monomial 2c) is in fact a digital signal n constituted by a small integer, while DP is a binary signal whose the value 1, by definition, illustrates a 25 exceeds the sensitivity threshold that is accumulated in the CO signal. Set 8 of FIGS. 17 and 18 consists of unilinear or one-dimensional type STN modules, of the bilinear or two-dimensional type and possibly of the tri-linear or three-dimensional type, the unilinear modules processing the single parameter CO, the bilinear modules processing two parameters, such as the x and y coordinates of the pixels according to the columns and rows of the matrix table of the pixels or radar signals, and finally the tri-linear modules processing three parameters, for example three parameters defining a color, and in the latter case, these three parameters are advantageously the luminosity L, the hue T and the saturation S. An embodiment of a one-dimensional module is described and illustrated in the publication WO-01/63557 having the same inventor, while an embodiment of bilinear and trilinear modules is described and illustrated in the French patent application N 01-02539 of the same inventor, depot dared on 23 February 2001 and entitled "Method and device for locating an object by its shape, size and / or orientation", published under No. FR-2821459 and in particular on page 6, lines 18 to 23 with regard to a bilinear module and on page 7, lines 1 to 7 with regard to a trilinear module, a bilinear or tri-linear module being, according to this patent application constituted by two, respectively three, unilinear modules of which the two, respectively three, output signals are applied to the inputs of an AND gate whose output constitutes the output of the pseudo-module

  bilinear, respectively trilinear.

  It is possible to use such types of unilinear, bilinear or trilinear modules to form unit 8 of FIGS. 17 and 18; however, it is best to implement modes of

  preferred embodiments described hereinafter with reference to FIGS. 20 and 23a-23b respectively concerning a unilinear module (denoted STN (1) and symbolized in FIG. 19) and a bilinear module (denoted STN (2) and symbolized in FIG. 22) , it being understood that a tri-linear module (denoted STN (3) and not illustrated) derives (as explained hereinafter) very simply, by some additions, from a bilinear module.

  Referring to FIG. 20, concerning the case of a single parameter, it can be seen that a unilinear or one-dimensional STN (1) module essentially comprises three units, namely a CH histogram calculation unit, a classification unit

  CL and a retro-annotation unit RA.

  The histogram calculation unit CH comprises several subunits and first a digital analysis memory 100, for example a DRAM or a SDRAM having an address number equal to the number of possible levels chosen for the values. of the input parameter DATA (A) namely CO and a word width representative of the number of pixels per frame (or image), namely for example 18 bits for an image of 256,000 pixels. This memory 100 has three inputs, namely a data entry 100a DATA IN for the data, a write input 100b WR and a

  100c ADRESS entry for addresses.

  The data input 100a is fed from a data input multiplexer MUX 106 which receives, on the one hand, at 106a, the output signal 107 'of an adder or incrementer S 107 for enabling the data input. incrementing and, on the other hand, at 106b, a signal "0", if the initialization signal INIT applied to the input

  106c of the multiplexer 106 is "1".

  The write input 100b of the memory 100 receives the output 15 of an OR gate 113 whose two inputs respectively receive the initialization signal INIT (this signal also acting as indicated above on the multiplexer 106) and the signal writing order

WRITE.

  As for the ADRESS input 100c of the memory 100, it receives the output signal 105s of a MUX address multiplexer 105 which has three inputs, namely an input 105a which receives the input parameter CO from the spatiotemporal smoothing unit 3 of FIGS. 17 and 18, an input 105b which receives the COUNTER signal from a counter (not shown) and an input 105c which receives the output of an energized adder / subtractor 108, of a on the other hand, at 108a, by the COUNTER counter signal and, on the other hand, at 108b, by the output signal 1045s of the POSRMAX register 1045 (which will be specified hereinafter) through an AND gate 109 with two inputs, one receiving the aforementioned signal 1045s and the other the signal ARME, this gate transmitting the signal 1045s only if the signal ARME is actually present, but forcing the value "0" in the opposite case; in addition, the adder / subtractor 108 is controlled, on its input 108c, by a sense signal, denoted SENS, consisting either of the signal "0" or of the signal "1", namely for example "0" to order an addition and "1"

to order a subtraction.

  Finally, the above-mentioned adder S 107 receives, at its first input 107a, a signal 100s from the output OUT 100d of the memory 100 and, on its second input 107b, a validation signal 102s coming from the retro-annotation RA as specified below; the output 107c of the adder S 107 outputs a signal 107s which is equal to the output of the analysis memory 100 (100s signal) if the enable signal 102s is equal to "0", but equal to said increased output of 1 if the validation signal 102s is equal to "1"; the output 107c of the adder 107 is connected to the input 106a of the multiplexer 106, as indicated above, and to the input P, namely P4, of a comparator 140, which compares this signal 107s with the signal of 1044s output of the RMAX register 1044, received on its input Q, to deduce from this comparison (if P> Q) a 140s signal for inserting POSRMAX (the position of RMAX) in the register 1045 through the

  AND gate 141 under the control of the WRITE signal.

  The histogram calculation unit CH of the STN module (1), which has just been described, finally outputs three output signals,

  namely 100s, debited by the memory 100, 105s, output by the multiplexer 105, and 107s, output by the adder 107, to the CL classification unit and receives the signal 1045s, from the register 1045 of this unit; furthermore the CH unit receives the signal 102s from the retro-annotation unit RA.

  As for the CL classification unit, which constitutes a

  passive classifier, it essentially comprises an operating sub-unit 101, a classification update sub-unit 103 and a storage sub-unit 104 having a plurality of registers.

  The operating sub-unit 101 acting as a classifier comprises two comparators 110 and 111 which receive, on their input P (P1, P2), the signal CO which constitutes the input data and, on their input Q (Q1, Q2 ), respectively the value of the CO classification terminal A and the CO classification terminal B, from the respective registers 1042 and 1043, discussed below. Each comparator 110 and 111 compares the values received at P with the values received at Q. If the value received at P1 or P2 is greater than that received at Q1 or Q2 (or possibly equal for P1 and Q1) respectively, each comparator sends a signal to an input a or b (after inversion for the latter) of an AND gate 112 which also receives, on its input 112c, from the temporal smoothing unit 3 of FIGS. 17 and 18, the signal DP indicating a exceeding by its value 1, or not exceeding by its value 0, a threshold in this unit 3. The output signal 101s of the AND gate 112, and therefore of the subunit 101, which is present only if simultaneously DP = 1 (overflow in the unit 114 of Figure 9), P1> Q1 and P2 <Q2, is applied to a bus 114, common to several STN modules. The classification update subunit 103 of the classification unit CL comprises a selector 1152, of which three inputs a, b and c respectively receive the outputs of the registers 1044, 1041 and 1046 of the subunit 104, know RMAX, NBPTS and 20 THRESHOLD, which will be mentioned below. The signal of the exit

  1152s of the selector 1152 is applied Q, denoted Q3, of a comparator 1151, which therefore receives either RMAX, NBPTS or THRESHOLD, as a function of a Choice signal received on the input e of 1152. The other input comparator 1151, denoted P3, is connected to the output 100d of the memory 100 to receive the signal 100s.

  The comparator 1151 checks whether the value of the signal on the input P3 is greater than or equal to the value of the signal on its input Q3 (P> Q). If this is the case, an output signal 1151s is applied to the inputs a of two logic controllers 1063 and 1064 which, moreover, receive on their input b the end of operation signal END and on their input c (after inversion for the second), the signal SENS (that also received by the adder / subtractor 108 on its input 108c); a fourth input d of each of the logic controllers 1063 and 1064 is adapted to receive an ARME arming signal enabling the task to be performed as long as the function represented by the signal 1151s received on the input a is true (P> Q), i.e. P3> Q3. The outputs e of the logic controllers 1063 and 1064 transmit signals 1063s and 1064s respectively which validate in In (enable) the registers 1042 and 1043 for the terminals Terminal A and Terminal B respectively, as explained hereinafter. The third subunit 104 of the classification unit CL comprises several registers, namely first a register 1044 for the maximum value RMAX, a register 1045 indicating the relative position POSRMAX of this maximum value, a register _0 1041 for the number of NBPTS points and finally a register 1046 of the THRESHOLD, the registers 1044, 1041 and 1046 output in the multiplexer 1152, as mentioned above, while the register 1045 outputs its output signal 1045s on the input 108b of the adder / subtractor 108 of the histogram calculation unit CH, when the AND gate 109 of the histogram calculation unit CH receives the signal ARME. Subunit 104 also includes two other aforementioned registers, namely 1042 for terminal A and 1043 for terminal B; each of these two registers receives, on its input En (Enable or validation or authorization), the validation signal 1063s, 1064s outputted by the output of the logic automaton 1063, 1064 respectively, and on its input a, the output signal

  s of the multiplexer 105 of the histogram calculation unit CH.

  The retro-annotation unit RA, the inputs of which are connected to the bus 114 common to several STN type units, essentially comprises a battery 102 of registers 102r denoted 1, 2, 3 ..., m, which respectively receive the values in1, in2, in3 ... inm of the output signals 101s of the subunits 101 of the

  different STN units connected to the bus 114.

  For each of the registers 102r, the unit RA performs the comparison between the input value, such as inm with the value contained in the corresponding register, such that m, and transmits to the input 107b of the adder 107 of the histogram calculation unit CH, a validation signal 102s equal to 1 each time there is equality between the value contained in a register 102r 35 and the corresponding input signal, but a signal 102s equal to 0 otherwise, by applying the Boolean formula signal 102 = (in1 + Reg-) (in2 + Reg2) .... (inm + i-egm) (in1 + ..... in2 ... + inm )

  Reg designating the previous contents of the register 102r.

  Preferably, the one-dimensional unit STN (1) of FIG.

  may also include registers 1047 and 1048 for storing the position of the average value representative of the input data CO; this POSMOY average position corresponds in a way to the center of gravity of the values of this CO data item.

  An entry, denoted a, of the register 1047, written POSMOYo, receives the signal 105s output by the multiplexer 105 of the histogram calculation unit CH, while the other input, denoted b, of the register 1047 receives, after inverting, the output signal 1153s of a comparator 1153 whose input Q5 receives the signal NBPTS (selected by the signal Choice from the signals received on the inputs a, b and c of the multiplexer 1152), after a division by two; The other input P5 of the comparator 1153 receives the output signal 1155s from a register 1155 which accumulates from receipt of an initialization signal INIT until receipt of an end of calculation signal END of the output at c of an adder 1156, the input of which receives the output signal 100s from the memory 100 and whose input b receives the digital signal 1155s of output from the

  register 1155, o cumulated 100s signals.

  The comparator 1153, if P> Q (actually P5> Q5), delivers an output signal 1153s which is applied, after inversion, to the input a of the first register POSMOYo 1047. The output signal 1047s of the register POSMOYo 1047 is received on the data input a of a second additional register POSMOY1 1048 until the arrival of an END signal on another input b of this register 1048. The register POSMOYo 1047 determines the current average position (POSMOY ), also referred to as the centroid, of NBPTS resulting from the comparison in the comparator 1153 of NBPTS / 2 and the accumulated value in the signal 1155s, while the register POSMOY1 1048 stores the average position (or the centroid) just anterior which is refreshed by the 1047s signal output by the

register POSMOYo 1047.

  Finally, note that the registers 1047, 1048, 1042 and 1043

  each comprise an additional input denoted by> 5 fed by a CLOCK-PIXEL signal. All these registers therefore operate in synchronous mode.

  We will now expose the operation of the module

  single-linear STN (1) of FIG.

  With regard to the operation of the unilinear block lo STN (1) illustrated in FIG. 20, this same, as well as the operation of the bilinear module STN (2) illustrated in FIG. 23a-23b and also that of the module of the type For each of the consecutive sequences, the trilinear circuit, not illustrated, comprises three successive phases or cycles, namely an initialization cycle, a calculation cycle and an update cycle of the registers storing the results. Now limiting itself to the case of the STN module (1) of the monolinear input figure, that is to say constituted by a single variable, which is in the particular case CO, the three aforementioned cycles are the following: Initialization cycle At the beginning of this cycle, the STN module (1) receives an INIT signal from an external sequencer (not shown). This signal INIT, a) forces the registers POSRMAX 1045, RMAX 1044, NBPTS 1041 and reg 1155 to 0; b) starts a counter COUNTER which scrolls the successive values 0 to M (namely the maximum size of the memory); c) forces at "0", via the multiplexer MUX 106 (which receives "0" on its input 106b and INIT on its input 106c), the input 100a DATA IN of the memory 100; d) places the memory 100 in write mode via the OR gate 113 through the input 100b WR of this memory 100 and, e) activates the output signal 105s of the multiplexer MUX 105 to impart, to the address of the memory 100, by its input 100c, the successive values of the counter COUNTER (not illustrated)

  via the multiplexer MUX 105.

  2) Calculation cycle Once the INIT signal has been stopped (ie at least after the end of the running time of the counter COUNTER, not illustrated, from 0 to M), the signal WRITE starts the calculation cycle which comprises for each cycle pixel (CLOCK-PIXEL): a) the accumulation of the value of CO in the memory 100 each time the signal of VALIDATION 102s is true, the latter being transmitted, as explained in relation with the structure of the module STN (1) by the output OUT of the unit 102, the accumulated values of CO in the adder 107 being transmitted to the memory 100 (at its input 100a) through the multiplexer 106 b) the comparison in the comparator 140, if signal 102 is true, from the value RMAX of the cofheater 1044, to the value of the signal 107s coming out of the adder 107; if the output of the comparator 140 is true, the value of the signal 107s, which constitutes a new value of RMAX, (since it is greater than that previously entered in the register 1044 as shown by the comparison in the comparator 140), is inscribed in FIG. register RMAX 1044 under the effect of the command from the AND gate 141 which receives, on the one hand, this output of the comparator 140 and, on the other hand, the signal WRITE, which carries out the inscription of the the value of the signal 107s, in the register RMAX 1044, while the value of the signal 105s, namely the position or address corresponding to this new value of RMAX, is written in the register POSRMAX 1045, the input of this register 1044 being also connected, as the input of the POSRMAX register at the output of the AND gate 141; at the same time the value entered in the NBPTS register 1041 is incremented by 35 "1"; c) the calculation of the output 101s of classification, of the value CO, in true value, whenever this value CO is

  between terminals A and B of unit 101.

  3) Update cycle of the result registers This last cycle starts when the signal WRITE becomes null (equal to 0) and the signal END becomes active on the multiplexer MUX 105, while the signal ARME is equal to "0" ; this register update cycle is constituted by two successive half-cycles. 3.1) The first half-cycle, which calculates POSMOY, comprises the following successive operations: a) the value of POSMOY0 of the register 1047 is transferred into the register POSMOY1 1048; b) the command Choice, applied to the selector 1052, forces the output d of this register to the value NBPTS / 2; c) the value of the counter COUNTER is transmitted by the multiplexer 105, because of the END command also arriving on this multiplexer and hence addresses the memory 100, by its input 100c, to this value; d) the scanning by the COUNTER signal makes all the stored values 100s that are accumulated in register reg 1155 come out, thanks to its initial value equal to 0 and to the summation made by the adder 1156 which receives, in addition to the signal 100s , the 1155s output of register 1155; e) as long as the value of the output signal 1155s of the register 1155 is smaller than or equal to NBPTS / 2, the comparator 1153, which compares these two values, validates by its output 1153s the writing of the value of the signal 105s in the register

POSMOYo 1047.

  3.2) The second half cycle calculates the classification terminals, namely Terminal A and Terminal B, from the POSRMAX maximum position. This second half-cycle implements the following operations: f) the counter COUNTER is reset to 0; g) the signal ARME is validated and therefore the value POSRMAX, constituting the output signal 1045s of the register 1045, is transmitted to the subtracter-adder 108 by the gate AND sh) the signal SENS command, in the adder / subtractor 108, the subtracting POSRMAX-COUNTER (the first value arriving at 108b and the second at 108a), the result of this subtraction finally constituting the output signal s of the multiplexer 105; i) the counter increments the COUNTER signal by one unit and this signal is valid as long as the output 100d of the memory 100 outputs a signal 100s greater than the output signal of the multiplexer 1152 (output d), the comparison being made in the comparator 1151 which outputs an output signal 1151s to the logic controllers 1063 and 1064; j) the validation logic function 1063 is true and validates the writing of the value of the output signal 105s into the terminal register A 1042 in which it arrives at the input a; k) as soon as the output signal 1151s is no longer valid, the process of the aforesaid operations c, d, e is interrupted and the value of terminal A remains frozen at the last value entered in the register 1042; 1) the counter (not shown) delivering the COUNTER signal is reset to 0; m) the signal SENS forces the addition of POSRMAX + COUNTER in the adder / subtracter 108, the result of this addition finally constituting the output signal 105s n) the signal ARME is validated o) the counter increments by one unit the a COUNTER signal which, in turn, increments the ADRESS 100c input of the memory 100 by one unit as long as the output 100d of this memory 100 outputs a signal 100s greater than the output signal of the multiplexer 1152 (output d), the comparison being made in comparator 1151; p) the logic validation function delivered by the unit 1064, namely the signal 1064s, is true, which validates the writing of the value of the signal 105s in the register B bound 1043; and q) as soon as the output signal 1151s of the comparator 1151 is no longer valid, the process of the operations i, j, k is interrupted and the value of the terminal B remains frozen at the last value

entered in the register 1043.

  These three cycles being completed, an external sequencer (no

  shown) can restart a new set of the aforementioned three cycles.

  The histogram of the values of the input parameter CO (which is the output parameter of the spatio-temporal smoothing set 3 of FIGS. 1 and 2 and which is detailed in FIG. 9) can comprise one or more peaks. FIG. 21a shows a histogram of CO parameter values with a single peak P, while FIG. 21b shows a histogram with two peaks.

P1 and P2.

  Referring to FIG. 21a, it can be seen that, in the case where the histogram of the CO values has only a peak P, it is possible to define on the histogram, which is plotted with the values of CO on the abscissa. and the frequency Q of these values on the ordinate, first of all RMAX, namely the ordinate of the peak or the peak P, then, starting from the position of RMAX, that of RMAX / 2, that is to say say half of RMAX, and finally POSRMAX, which is the abscissa of peak P. The values of RMAX, on the one hand, and POSRMAX, on the other, are, as previously explained, stored in the corresponding registers that is, 1044 and 1045. RMAX / 2 defines the terminals A and B shown in Fig. 21a, these two terminals being stored respectively in registers 1042 and 1043; these terminals are also applied to the comparators 110 and 111 to be compared with the input parameter CO. The output signal 101s of the set 101 of the aforementioned comparators 110 and 111 and the AND gate 112 is illustrated at the bottom of FIG. 21a and it is seen that it comprises a slot (positive) between the terminal A and the terminal B. Figure 21b shows the case where the histogram of CO values has two peaks P1 and P2 (instead of the single peak P of Figure 21a). At the peak P2 corresponds, as in the case of Figure 21a, RMAX, on the one hand, and POSRMAX, on the other hand. Furthermore, from RMAX, RMAX / 2 is defined as in the case of FIG. 21. The difference with FIG. 21a is that the peak P1 also defines a maximum value which is referenced RMAX '. The value of RMAX / 2, instead of defining only two terminals, namely Terminal A and Terminal B, defines, in connection with FIG. 21b, two pairs of terminals Terminal A, Terminal B and Terminal A ', Terminal B' which are manifested by (positive) slots of the signal 101s (a) at the bottom of this figure 21b. Since POSRMAX, which corresponds to the highest peak P2, has been determined, it is possible within the scope of the invention to eliminate from the signal 101s (a) the first slot corresponding to the smaller amplitude peak P1. and, finally, a signal 101s (b) is obtained which comprises only the slot corresponding to P2 and which is delimited by the terminals A and B defined by RMAX / 2, as in the case of FIG. 21a. It is finally the signal 101s, in its variant b, which is transmitted to the bus 114 in the case of two or more peaks, this signal, which has the same shape as the signal 101s of FIG. 21a, being defined by the terminals A and B corresponding to RMAX P2 peak

The highest.

  After describing the structure, operation and results of an STN module (1) (FIGS. 19, 20, 21a and 21b), we will

  hereinafter give analogous explanations with respect to an STN module (2) diagrammatically shown in FIG. 23 and illustrated in detail in FIG.

Figure 23a-23b.

  The STN module (2) detailed in FIG. 23a-23b comprises, like the STN module (1), essentially three units, namely a CH histogram calculation unit, a CL classification unit and an RA retroannotation unit. , and delivers on a bus 114 common to several STN modules; furthermore, advantageously, as in the case of the STN module (1), the STN module (2) may further comprise registers provided in the CL classification unit and discussed below to store the position of the average value representative of the input data (there are two input parameters in the case of the STN module (2)), 1156 units,

1155 and 1153.

  The histogram calculation unit CH of the register STN (2) of FIG. 23a is distinguished from the corresponding unit CH of the module STN (1) of FIG. 20 by the fact that, with regard to the input 105a of the multiplexer 105, not the input of a single DATA input signal (A), namely CO, but of two associated input signals 1o DATA (Ay) and DATA (Ax), for example the two cartesian coordinates y and x extracted, by the unit 9 of FIG. 1 or FIG. 2, from the input signal 7 (after defocusing) or 7 'respectively (before processing in the filter 4 carrying out an operation similar to a defocus). To account for the processing of two input parameters instead of one, the CH unit of Fig. 23a has, in addition to the STN module (1) of Fig. 20, a shift shift unit 130. on the left (multiplication) which receives the signal END (,) on the one hand, and the signal MOYy on the other hand, the association of these two signals applying an offset equal to the number of bits of the signal

  DATA (Ax) at the input signal COUNTER to output a signal 130s on the input 108a of the adder / subtractor 108, whereas, in the case of an STN module (1) (FIG. 20), the signal COUNTER is applied directly without offset to the input 108a of the adder / subtracter 108.

  In module STN (2) (Figure 23a-23b), the unit of

  CL classification comprises a number of duplications with respect to the corresponding unit of the STN module (1) of Figure 20.

  More particularly, the storage sub-unit 104 includes, not a pair of registers 1042 and 1043 for the terminal A and the terminal B respectively, but two pairs of registers, namely 1042 (1) and 1043 (1) for the Ax terminal (relative to the input x parameter) and the Bx terminal (also relative to the x parameter), on the one hand, and a pair of registers 1042 (2) and 1043 (2) for the Ay terminal (relating to the 35 input parameter y) and for the By terminal (also relative to the input parameter y), these two pairs of registers receiving, as input signal, the signal 105s output by the multiplexer 105.

  Because of the duplication of register pair 1042 and 1043 for terminals A and B respectively, the two PLCs 1063 and 1064 of the STN module (1) of FIG. 20 are also split in the case of an STN module (FIG. 2) and in FIG. 23b, two pairs of such automata are illustrated, namely 1063 (1) and 1064 (1) corresponding to the input parameter x and 1063 (2) and 1064 (2) 1o corresponding to the parameters of entry there.

  Similarly, the operating subunit 101 of FIG. 20 is split in the case of an STN module (2) of an assembly 101b, in subunits 101 (1) and 101 (2), playing a similar role. to that of the subunit 101, each of these subunits 101 (1) and 101 (2) delivering a signal, respectively 101 (1) s and 101 (2) s, analogous to the single signal 101 (s) delivered by the STN module (1) of FIG. 20; the two signals 101 (1) s and 101 (2) s are debited on the inputs a and b of the AND gate 131 which delivers an output signal 101s only in the case where the two signals 101 (1) s and 101 (2) s arrive simultaneously on the two inputs of this door

  131; it is this signal 101s which is output in the bus 114.

  To complete the description of Figure 23a-23b, it is reported

  that it comprises, in addition to FIG. 20 concerning a unilinear module STN (1), anticipation means 25 which consist of an anticipation block 1049 (FIG.

  receives, on its two inputs, the outputs of the registers POSMOYo1047 equal to POSMOYxy relative to the two input parameters x and y, and POSMOY1 1048 and which determines, in response, the differentials Ax and Ay representative of the value of the two registers 1047 and 1048.

  The registers POSMOY0 and POSMOY1 flow in a calculation module 1049 which calculates the variation of each POSMOY between two consecutive frames, as well for DATA (Ay), DATA (AX). The variation of POSMOY for the parameter y is denoted by Ay whereas that for the parameter x is denoted Ax; Ay and Ax calculated in the calculation module 1049 are respectively debited on two subtracters 145 and 146, which receive these values on their input b while their input a respectively receives DATA (Ay) and DATA (Ax). The outputs of the 5 subtractors 145 and 146, which calculate respectively DATA (Ay) Ay and DATA (Ax) - Ax, of o anticipation based on a linear variation of DATA (Ay) and DATA (Ax), are debited concomitantly.

  in the unit 101b to output a signal 101s of classification.

  The output POSMOYxy of the register 1047 is applied to the input b 10 of a comparator 1050 (FIG. 23a) which determines the equality (or not) between POSMOYxy and the input parameters DATA (Ay, Ax); this comparator 1050 delivers an output signal 1050s, denoted BarZ1 (BarZ being the abbreviation of zone barycenter) and i representing the current value of the zone portion. Explanations of this additional set of 1049 and 1050 and the

  1050s signal will be provided below.

  The operation of the STN module (2) of FIGS. 22 and 23a23b is inspired by the operating mode of the STN module (1) of FIGS.

  Figures 19 and 20 with the differences below.

  Essentially the initialization cycle 1) of the STN module (1) is kept as it is for the STN module (2); on the other hand, cycle 2) calculation, the classification phase is split; finally, in the cycle 3) update of the result systems, the DATA (AX) and DATA (Ay) are processed successively. 1) Initialization cycle This actually comprises the arrival of the signal INIT, the operations a, b, c, d and e are detailed above in the case of

operation of an STN module (1).

  2) Calculation cycle The STN module (2) receives the input parameters DATA (Ay) and DATA (AX) on the input 105a of the multiplexer 105; it processes these data DATA (Ay) and DATA (Ax) in the same way as the unique data input DATA (A), namely CO, of the STN (1) of FIG. 20; in fact, in the two embodiments (STN (1) and STN (2) modules), the input parameter CO or the input parameters DATA (Ay) and DATA (AX) are sent by the multiplexer 105 , in

  as signal 105s, at the input 100c ADRESS of the memory 100.

  The AND gate 131 combines the output signals 101 (1) s, from the subunit 101 (1), and 101 (2) s, from the subunit 101 (2), to finally debit. , in case of simultaneous presence of the two signals, the global output signal 101s

applied to the bus 114.

  3) Update cycle of the result registers The data processing, DATA (Ay) implements, among 1o the duplicated blocks, those having the index (1), namely the registers 1042 (1) and 1043 ( 1) for the terminals, Terminal Ax and Terminal Bx, the logic controllers 1063 (1) and 1064 (1), and the operating sub-unit 101 (1). Once the processing of the input parameter x, namely DATA (Ax), has been completed, the unit 108a triggers the processing of this data item x to that of the second

  input data y, namely DATA (Ay), the latter being processed analogously, but using among the duplicated blocks those comprising the index (2), namely the registers 1042 (2) and 1043 (2 ), logic controllers 1063 (2) and 1064 (2) and operation subunit 101 (2).

  This last cycle starts when the WRITE signal becomes zero (equal to 0) and the signal END (1, 0) becomes active on the multiplexer MUX 105, while the signal ARME is equal to "0" this update cycle registers is constituted by: 3.1 the first third of the cycle calculates POSMOY with the operations a to e previously described, 3.2 the second third of cycle computes the classification terminals Borne A (x) and Borne B (x) from the portion DATA (Ax) with the operations a to I previously described, 3.3 the third third of the cycle calculates the classification boundaries

  Terminal A (y) and Terminal B (y) from the portion DATA (Ay).

  The unit 130 is activated by the combination of commands MOYy and END (1) and the operations a to I previously described.

are performed.

  FIG. 24 represents, in two dimensions along x and y, namely DATA (Ax) and DATA (Ay), two peak areas V1 and V2, in a manner similar to FIG. 21b illustrating, in one dimension according to x ie CO, a curve with two peaks P1 and P2; in FIG. 24, an example of the results obtained with a two-dimensional module STN (2) of the type illustrated in FIG. 23a-23b is illustrated. In the three-dimensional Cartesian coordinate system, we find according to the coordinates (x) and (y), DATA (AX) and DATA (Ay), and according to (z) the number or quantity Q of points having the lo coordinates (x) and (y). In relation to the highest peak V1, there is shown RMAX, POSRMAXx and POSRMAXy and the terminals, Terminal Ax, Terminal Ay, Terminal Bx, Terminal By; on the other hand, for simplicity, these values have not been represented for the peak V2. As in the case of FIG. 21b, the use of POSRMAX / 2 (not shown) makes it possible to eliminate the secondary peak V2 so that

  consider that the main peak V1. Figure 23c illustrates a partial variant of the STN module (2) of the

  FIG. 23a-23b, namely a variant of a portion of the calculation unit of the CH histogram shown in FIG. 23a, the difference being constituted by the addition, in the CH unit, of FIG. 23c, of a calculation block 120 delivering its output signal 120s on the input "1" of the multiplexer 106 and receiving on a first input the output signal 100s of the memory 100 and on a second input a coefficient Km , advantageously constituted by a power of 2, namely Km = 2m. This calculation block 120 performs on the signal 100s OUT of the memory 100 and the coefficient Km (applied on its two inputs) the function of

  transfer "120s" = (K-) x "100s">.

  The coefficient is an information retention parameter 30 plus m is large minus it is possible to modify the storage of the histogram in the memory 100; on the other hand, when m = 0, the block output 120 is equal to "0" because K = 2 = 1 and therefore Km1 = 0 and we end up in the case illustrated in FIG. 23a on which the input 1 of the multiplexer 106 is constantly equal to

  "0" during the initialization cycle.

  In the case where m = 0, the calculation cycle stored in the memory 100 is recalculated at each sequence; on the other hand, when m 5 is different from <"0", a percentage (more or less high depending on the value of m) of the storage of the memory 100 is kept

for the current calculation cycle.

  The assembly of FIG. 23c performs an improvement over that of FIG. 23; in fact, in the case of FIG. 23 23a, as explained above, in the first cycle, namely the initialization cycle (INIT = 1), the memory 100 is reset by the multiplexer 106, which forces to zero its input 100a for DATA IN; on the contrary, in the arrangement according to FIG. 23c, a memory effect is added to the STN block (2) by replacing the arrival of the "zero"> signal on the input 1 of the multiplexer 106 by the arrival of the signal of output 120s corrected by the factor Km 1 Km However, the implementation of said transfer function in block 120 makes it necessary to increase the length of the "words>" stored in the memory 100, and therefore to oversize it, in relation to in memory 100 of FIG. 23a; indeed if, by

  For example, the memory 100 of FIG. 23a was designed to store 10-bit words and if the coefficient Km (applied to the block 120 of FIG. 23c) can vary from 1 to 2 m, the capacity of the memory 100 of FIG. 23c must be sized to store words of 25 (10 + m) bits.

  The signal Km is outputted by a register 121 controlled by the API of the module STN (2), depending on the importance of

  the desired learning for this module.

  It should be noted that the addition of the block 120 and the register 121 illustrated in FIG. 23c for a two-dimensional module STN (2) can also be applied to a one-dimensional module

  STN (1) as shown in Figure 20.

  Finally, in addition to the single-linear and bilinear STN modules, the unit 8 of FIGS. 17 and 18 may comprise tri-linear STN modules, for example for the three parameters.

  representing a color, such as LTS (or RGB) parameters.

  A tri-linear module is analogous to the bi-linear module of FIG. 23a-23b (possibly with the improvement of FIG. 23c), except that: * on the one hand, the block 130 of the calculation unit of histograms CH of an STN register (3), instead of being able to impart a single offset, as in this module STN (2), is capable of controlling two offsets, namely for the second and third DATA (the three DATA being for example DATAL, DATAT and DATAs; on the other hand the units referenced (1) and (2) of FIG. 23a-23b are accompanied by a third unit similar to the units 101 (1) and 101 (2). , 1063 (1) and 1063 (2), 1064 (1) and 1064 (2), 1041 (1) and 1041 (2), 1042 (1) and 1042 (2); the calculation of POSMOY values and calculations; anticipation

  and classification are triples instead of doubles.

  The operation of a tri-linear STN module is analogous to that of the bi-linear module illustrated in FIG. 23a-23b except that the three parameters are successively processed in three referenced sets (1), (2). ) and (3) instead of only two sets (1) and (2), the three parameters such that L, T and S being successively processed as well as successively processed, in the bi-linear module, the parameters x and y. Thus, terminal AL, terminal BL, terminal AT, terminal BT, terminal As and terminal Bs are obtained; finally, while the overall output signal of the assembly 101 of FIG. 23a-23b is constituted by a signal 101s (applied to the bus 114) delivered by the output of an AND gate 131 with two inputs (receiving the outputs of FIG. 101 (1) and 101 (2)), the output signal of a three-dimensional module STN (3) is constituted by the signal delivered by the output of an AND gate (not illustrated) having three inputs for receiving the outputs units 101 (1), 101 (2) and additional unit 101 (3) (not shown) similar to units 100 (1) and

101 (2).

  Figure 25 illustrates a solution for improving the classification of multi-terminal DATA parameters (in any number) with anticipation. In this figure 25, we have taken

  most of the STN (2) devices of Figures 23a-23b or 23a-23c.

  FIG. 25 is identical to FIG. 23a-23b except the unit 101 which is replaced by the unit 101c constituted of a memory 118, equal in number of words to the memory 100 but in width of

  1 bit word, and by the addition of a multiplexer 144 with 2 inputs.

  This multiplexer 144 is controlled by an end operation signal END, which when it is equal to zero, outputs the result of the units 145 and 146 and when it is equal to one, outputs the signal 105s. . The multiplexer 144 outputs an output signal 144s to the ADRESS input of a memory 118 whose DATA IN input consists of the output of a comparator 1151 which receives at its input Q half of RMAX from its register and on its input P the output 105s of the memory 100; this comparator 1151 compares the input signal on its input P and that of its input Q and if P> Q, the comparator 1151 outputs a signal, equal to one, on the input DATA IN of the memory 118, and equal to 20 to zero in the opposite case. It is this signal on the output OUT of the memory 118 which then constitutes the signal 101s of output, with

  anticipation, applied to bus 114.

  The memory 118 of this module 101c is written during the update cycle of the result registers, signal END equal to one. For each of the addresses of the memory 100 and thus of the memory 118, through the action of the multiplexer 144, a comparison is made between the value of the memory 100 and a threshold defined by the command CHOICE, and defined by the value of the bit to be written in the memory 118; equal to 1 if the value of the memory 100 is greater than or equal to the zero threshold otherwise, the write command of the memory 118 being defined by the signal END. The classification with anticipation is carried out during the calculation cycle, through the action of the multiplexer 144, which transmits the result of the parameter DATA, modulated by anticipation, as the address of the memory 118, the OUT

  of this memory defines the classification value.

  Apart from this, the implementation of the improvement of the figure

  results in a similar, though improved, result to that achieved by the operation of the means illustrated in FIGS. 23a-23b, or 23c-23b.

  The appearance of two peaks during the same sequence as a result of the abrupt increase in the resolution at the beginning of this sequence, for example in phase t1 (FIG. 3) is illustrated in FIG. 21b (already considered). in the case of a one-dimensional histogram (peaks P1 and P2) and in FIG. 24 (also already considered) in the case of a two-dimensional histogram (peaks V1 and V2). These peaks correspond to two classes in the classification of the histogram of values of CO (FIG. 21b) or DATA (AX) and DATA (Ay) (FIG. 24). The one-dimensional module STN (1) of FIG. 20 or the two-dimensional module STN (2) of FIG. 23a-23b makes it possible to extract the dominant class represented by the peak P2 or

V1 respectively.

  The occurrence of several peaks or classes occurs, for example, when analyzing a scene observed, generally, with a MVT motion parameter, since there may be global or overall motion, for example. for example, a traveling or zooming effect of the lens with respect to the scene associated with particular movements, or an expansion from an expansion point (particularly in the case of an on-board camcorder).

  in a car traveling on a road).

  In the case now envisaged of at least two peaks appearing during the same sequence, it is advantageous to extract more than one class during this sequence in order to save processing time and preserve the momentum.

of these classes.

  A simplified solution consists of determining the dominant class (peak P2 or VI) in a first module STN during a first sequence, to eliminate by inhibition this class, to determine the new dominant class, and (after said inhibition) on a second STN module during the following sequence, to inhibit this new dominant class, to determine, on a third STN module during a third sequence, a new dominant class, and so on until the last class. Such an arrangement is illustrated in FIG. 26 (plate 12) comprising three STN modules, of unidimensional type STN (1), for example according to FIG. 20, processing the MVT motion and denoted ST1, ST2 and ST3; these modules all receive on their input a, the 0 parameter Z; the output signal MVT1 of ST1, representative of the first dominant class of MVT, is sent to time t, as an inhibition signal, on a second input b, of inhibition, ST2 established at time t2, while the output signal MVT2 of the module ST2, representative of the second dominant class of 15 MVTI, and the output signal MVT1 of the module ST1, representative of the first dominant class, are transmitted, as inhibition signals, on the second and third inhibition inputs b and c of ST3 which outputs the MVT3 output signal representative of the

* third dominant class of MVT.

  Such a process unfortunately requires a lot of time

  treatment, spread over several phases t0, tl, t2 (Figure 3).

  A solution for increasing the speed is illustrated in FIG. 27, in which the one-dimensional STN module, denoted ST'0, operating during the phase to extract all the classes of the MVT signal during the CALCULATION cycle and, during the cycle RESULT , extract one after the other, the different classes, in descending order of quantity, and rewrite them in the associated STN modules. The first dominant class found is transferred into a second one-dimensional module STN, denoted ST'1. The classified values of the histogram memory 100 are then counted to create a new value NBPTS transferred and set to zero, which makes it possible to determine a new dominant class in this second block ST'1, which outputs the signal MVT1, representative of the new dominant class of MVT. The transfer operations of the new classification result and the new registers 104 of the unit ST'0 are repeated, this time supplying the module STN, denoted ST'2, which outputs a new MVT2 dominant class of MVT, and thus immediately, until the histogram values 100 are exhausted in the memory 100. Thus in FIG. 27, there is illustrated a third unidirectional module ST'3 which outputs a signal MVT3 representative of a new class

dominant MVT.

  Another way to speed up the determination of several classes appearing in the same sequence is to locate these classes to determine levels of the amount of pixels to

from the maximum value.

  One can compare such a method, applied to a bilinear histogram DATA (Ax) and DATA (Ay) of the type illustrated in Figure 24 with at least two peaks V1, V2, when the resolution increases by 15 steps, a hilly topography with at least two vertices, covered with a sheet of water, when this level decreases to reveal at least two peaks for some degree of lowering of the water level. Indeed, when two or more vertices appear at the same time, by lowering the water level by a certain height, they must be classified by decreasing height. This illustrative comparison makes it possible to better understand the process of extracting successive classes which

will be explained below.

  For example, in a bilinear module STN (2) such as that of FIG. 23a-23b advantageously modified according to FIG. 25, the histogram comprising several peaks is stored in the memory 100, whereas the classifier 118 constitutes a memory at one bit (0 or 1), the value 1 corresponding to an overshoot by the value of the parameter, such as A, determined by TERMINAL Ax, TERMINAL 30 Bx, TERMINAL Ay, TERMINAL By, according to the diagram: Histogram memory 100 Classifier 118 address quantity associated class 0 x 0 1 x 0 2 x 0 3 x 0:. x 0

I A 1

  It can be seen that classifier 118 outputs "1" as signal 101s when the value A or B of the parameter exceeds the expected threshold, such as RMAX / 2, at the output of selection unit 1152 (FIG. Figures 20,

23b or 25).

  To extract several classes during a sequence, it is sufficient to order the RMAX by decreasing values in the register 1044, to memorize the thresholds A / 2, B / 2 of the classes and to provide a register for the number of classes whose thresholds have

have been memorized.

  The following scheme is thus carried out MO: memory of M: memory of the M2: classifier M3: memory Rc: the histogram 100 multi-class addresses of the thresholds of the registers ordered in classes of the RMAX number of classes Addresses Quantity Addresses RMAX Addresses Class Addresses Thresholds pt p rated pt Rc 0 x 0 i 1 1 A / 2 2 1 x 1 j 2 2 B / 2 2 x 2 3 x 3 i AII (1) j B j 0 j (2) nxnnn Cel Sorting and sorting RMAX decreasing value scheduling of the contents of the memory 100 to obtain the storage of the decreasing RMAX ordered addresses is achievable by a processing software which comprises three successive cycles of initialization, calculation and updating of the result registers. . 1. The initialization cycle consists of resetting all

  the memories Mo, M1, M2, M3 and the register Rc.

  2. The calculation cycle essentially consists in determining and memorizing in the memory M0 the values of the histogram

determined by the STN module.

  3. The updating cycle of the result registers, whose flowchart is illustrated in FIG. 28, comprises three successive sub-cycles A, B and C. It is during this update cycle of the registers of results (illustrated in Figure 28) that is carried out the extraction of classes in the case where several classes appear during the same sequence (constituted for example by one or

several frames or video images).

  The first sub-cycle A initially performs the sorting by

  decreasing values, RMAX contained in the memory M0 histograms and the registration in the memory M1, successive addresses of these decreasing values, as indicated by (M1) = Sort (M0) in Figure 28 (the memory M1 receives the result of the sorting of the memory M0).

  The second sub-cycle B searches for and defines the different classes of the histogram contained in the memory MO. It comprises several stages, namely successively i. A first initialization step 500 which sets, on the one hand, a pointer p which traverses the successive addresses, from 0 to n inclusive (via i and j), of the memory M1 and, on the other hand, on the other hand, the register Rc of the number of classes in which is registered the number of classes previously found (p = 0 and Rc = 0). ii. A second step in which a loop rolls successively all the values of p, from 0 to n inclusive, to execute, for each of these values, the three successive operations of the block 501 of FIG. 28, namely successively a) the determining a position Pt of the pointer equal to the contents of the sorting memory M1 at the address p (501a, FIG. 28), namely pt = (M1) p, that is to say the contents of the memory M1 at the pointer p; b) determining a value A equal to the contents of the memory M0 histogram address pt (501b Figure 28), namely A = (M0) pt; and c) recovering the values noted a, b and c (or a, b ...) adjacent to that of the class selected in the class memory M2 at the level or address pt, namely pt + 1 (a, b) for a single histogram (two positions), ptx + 1 and pty + 1 a, b ... for a bilinear histogram (eight positions), ptx + 1, pty + 1 and pt7 + 1 (a, b). .) for a trilinear histogram (26 positions), as summarized in 501c in Fig. 28; indeed, in the case of one-dimensional processing, there are two locations pt + 1, on either side of pt; in the case of a two-dimensional processing, there are eight locations pt + 1, namely on the peripheral squares of the large square area constituted by 3 x 3 juxtaposed squares whose square pt constitutes the center; in the case of a three-dimensional processing, there are 26 locations pt + 1, namely in the peripheral cubes of the large cubic volume constituted by 3 x 3 x 3 cubes juxtaposed

whose cube pt occupies the center.

  i. A third test step 502 to determine if there is no ("YES") or there are adjacent values (the absence of these

  values corresponds to the case where a = b = 0).

  ii. A fourth step with two possible treatments First case: the test 502 revealed that there are no ("YES") adjacent values, such as a, b ...; in this case a controller 504 is activated to perform the following successive operations: a) the register Rc of the number of classes is incremented by "1", which is classically noted Rc ++ (operation 504a, FIG. 28); b) the associated threshold A / 2 is stored in the memory M3 (FIG. 28) at the address defined by the value of the register Rc (operation 504b, FIG. 28); and c) the class defined in the register Rc (504c, FIG. 28). ) i0 is entered in the class memory M2, at the address

  defined by the position pt of the pointer.

  Second case: the test 502 revealed that there is at least one adjacent value: in this case a controller 503 is activated to execute the following successive operations: a) the logic operations 503a, 503b and 503c select the lowest value (non-zero) class among the different values of classes (a or b) adjacent a, bb) this value is written in the memory M2 of classes 20 at the address defined by the value pt (501a, FIG. 28), in distinguishing the case of the registration of a non-zero "a" from that of a non-zero "b" (operation 503d or 503e respectively). At the end of the subcycle B, the pointer parameter p is incremented by one unit (p ++ of operation 510) regardless of the

  test result 502, 503a, 503b, 503c.

  The third sub-cycle C updates the modules

  STN required for the detection of at least two classes found during the same sequence, as illustrated in FIG. 28, this subcycle having several steps.

  i. The first step involves an operation 505 that

  "1" the value of the current class register CI-Actu.

  (The following steps comprise, nested one inside the other, two loops, namely 508 and 509, the latter 35 housed in the first) 79. The second step comprises the first operation of the loop 508, namely the flow of all ClActu classes found, from 1 to the final class defined by the register

Rc of the number of classes.

  iii. The third step is performed by a controller 507 which - assigns to a new STN module a sequence number equal to the value of the current class CI-Actu, - determines the decision threshold corresponding to the content of the memory M3 to the address defined by the current class 10 CIActu, and - initializes to "0" the pointer p. iiii. The fourth step is performed by the loop 509 whose first operation consists of performing a test (operation 509a) to determine whether the content that has just been transcribed in the memory M2, (operations 503d, 503e, 504c, FIG. 28)

  at the address p of the pointer, is equal or not to the value of ClActu.

  if this content is different from the IC-Actu value, the controller resets the contents of the memory 118 of the previously selected STN module (operation 509b), but if the content is equal to the value of CI -Actu, the test 509c determines whether the content of the histogram reading M0 at the address p is greater than the value of the threshold and, in this case, the contents of the memory 118 1 bit of the STN module allocated is set to "<1" at address p; otherwise, this content is set to "0" (operations 509d and 509b respectively). iv The final step is to increment the address of the pointer p (operation 509e) by "1" and, when this address exceeds n (operation 509f), to increment by "1" CI-Actu (operation

  511); then loop 508 can start again.

  Finally, the new value of CI-Actu is injected at the entrance of

  loop 508, i.e., operator 506, in place of the value prior to IC-Actu.

  It is the assembly of FIG. 27 (discussed above) which makes it possible to implement the software with three sub-cycles A, B and C. In fact, this software implements a minimum number of treatments equal to n [log (n) + Rc + 1]; if n is high, the number of treatments 5 may be very large, therefore relatively long; in this case, it is better to do the class extraction differently,

as explained below.

  To overcome this difficulty, it is possible to implement an electronic assembly for updating the sorting of classes _o during the calculation cycle. In this case, memories M0 and M1

  are ready at the end of this calculation phase and the software of the results display cycle only includes sub-cycles B and C and requires only a reduced duration, which is very advantageous. The minimum number of treatments becomes equal to 15 n (Rc + 1).

  Figure 28a shows the flowchart of Figure 28 applied as an integrated API for maximum optimization in execution speed. Subcycle A is integrated in the calculation phase

and will be explained later.

  The flowchart starts at the beginning of the updating cycle of the result registers, directly in the sub-cycle B, by searching and defining the different classes of the histogram M0, associated with its sorting M1, contained in the unit 600 and obtained by the previous calculation phase. This sub-cycle B has the same 25 steps as in FIG. 28 except for the following details: * in the initialization step 500, a BUSY flag is validated, * the operations a) and b) of the block 501 of FIG. 28 are replaced by * the address switching command Sel.pp is disabled, (514, FIG. 28a) allowing the access of the unit 600 by its address decoding 601 in rank p, * the determination of a position pt of the pointer equal to the contents of the sort memory M1 at the address p, that is (POSRMAX) p (501a ', figure 28a), * the determination of a value A equal to the contents of the memory M0 d histogram at address pt, that is (RMAX) p (501b ', figure 28a), and * change of address switch Sel. pp is validated in order to access the different memories with the same fields

  address (p equivalent to DATA (A)).

  The third sub-cycle C updates the memory M2, serving as classifier in the calculation cycle. This subcycle C has the same steps as in FIG. 28 except where step 509b is canceled, and steps 509c and 509d in FIG. 28 are replaced by steps 509c 'and 509d' of FIG.

figure 28a.

  The test 509c 'determines whether the RMAX content of the reading of the histogram 600 through the multiplexer 105 at the address p is greater than the value of the threshold and in this case an operation (509d', FIG. 28a) reading of the memory M2 is made through the multiplexer 105, a flag Val is hung

  and everything is rewritten at the same address.

  At the end of the sub-cycle C, the transcribed class Rc classes, the sel.pp switch control is disabled (512, FIG. 28a) and the

  BUSY flag canceled (513, Figure 28a).

  It may also be advantageous to know the number of

  pixels belonging to each of the classes found, in this case an M4 memory is dedicated to the NBPTS values of each of the 25 classes described.

  By taking up the flowchart of FIG. 28a and completing it, in the subcycle C, the NBPTS value corresponding to the memory position M4 of the CI-Actu address is updated by adding the following two steps: * an initialization from the memory M4 to the CI-Actu address in the step 507, FIG. 28a, * an accumulation of the valid RMAX values in this memory position M4 at the output YES of the test 509'c by the function

  (M4) cI-Actu = (M4) cI-Actu + (RMAX) p.

  The enumeration (NBPTS) of the classified elements also makes it possible to calculate the centroid (POSMOY) of each of the classes by the addition of a subcycle D at the end of the sub-cycle C. By resuming the flowchart of Figure 28a and by 5 supplementing, in this sub-cycle D, are used two additional memories, a temporary memory MT calculation and a memory M5 backup of the results POSMOY related to each

  classes defined previously.

  This sub-cycle D starts with a zero initialization of the memories MT and MS, and the pointer p. - Follows a loop of calculation of the POSMOY values for all

the values of p from 0 to n included.

  - The first operation of this loop is to read the value

  class C1 of the memory M2 to the address p and to check whether the associated flag Val is present or not.

  [Val, Cl = (M2) p] - If this flag is present, the following automaton checks that the

  The content of the temporary memory MT at the address Cl is less than the content divided by two from the memory M4 to the address Cl.

  [(MT) cl <(M4) cl / 2] - If the previous test is checked, then the RMAX histogram value is added to the p address in the MT memory at

the address Cl.

  [(MT) cl = (MT) cl + (RMAX) p] and there is backup of the pointer p in the memory POSMOY

M5 at Cl.

  [(M5) ci = P] - the final step is to increment the address of the pointer p by "1" and restart the loop as long as this address

  is less than or equal to the value n.

  In the case of a multilinear histogram calculation, the pointer p represents the various fields of the data DATA. The previous calculation loop defines the POSMOY field portion by the leftmost field portion of the p value. It is therefore necessary to repeat the loop processing as many times as the number of fields defining the data DATA (A), and by swapping the fields of the value p.

  For example, for a bi-linear data DATA (Ay, Ax), it is necessary to scroll the loop twice, a first time with p running from 0 to n, the result corresponding to the part y POSMOY (y), then the second time by switching the two fields of p which scrolls from 0 to n, the result then corresponds to the part x POSMOY (x). The combination of these two fields corresponds to the 1o centroid POSMOY (x, y).

  We have just described a method and an associated device allowing to extract several classes and their centers of gravity

  associated from a multilinear parameter.

  In FIGS. 29a and 29b, one side (FIG. 29a, which summarizes FIG. 23a-23b) is arranged side by side, the embodiment without class sorting during the calculation cycle according to the described method. above and, on the other hand (Figure 29b), the mode of

  realization with class sorting during the calculation cycle.

  This version of FIG. 29b with integrated sorting corresponds to the assembly, in a single unit 600, of memories M0 and M1.

  this unit M0-1 containing opposite the RMAX and the POSRMAX for the values of 0 to p (inclusive) of the pointer, the ordering being done in descending order of RMAX starting from the position 0 to the position n (as indicated on the two previous tables).

  This unit 600 sorts the RMAXs in decreasing order, and thus also POSRMAX, and the API 602 controls an algorithm that reads directly at the address p (between 0 and n) decoded by the unit 601, the maximum value. of the histogram 30 (RMAXn), (A = (RMAX) p operation 501b ', FIG. 28a) and its position

  actual (POSRMAXn) (pt = (POSRMAX) p operation 501a ', Figure 28a).

  The RMAX content is also accessible by its POSRMAX position by means of a multiplexer 105, controlled by a signal Sel.pp, which directs the address value ADRES-p in the unit 600, the output signal OUT of said unit is then directed on the input LECT-p of the unit 602 via a multiplexer 654 controlled by the signal Sel.pp from the unit 602. This operation is performed during the test (operation 509c ', figure 28a) of the value

  RMAX at position p with a predefined threshold value.

  In the calculation cycle, the memory M2 acts as classifier multiclasses. the signal Sel-pp being then disabled, the multiplexer 105 transmits the signal DATA (A) as an address to the memory M2 which outputs a signal 651s accompanied by a signal Val, active in the case of a defined class. This signal 651s passes through 1o of a demultiplexer 653 which is validated on its input En by the signal Val coming from the memory M2. This unit 653 outputs signals C1 to Clk representative of the class of membership of the

DATA input signal (A).

  The assembly 603 of the units 600 and 601 is illustrated in greater detail in FIG.

  FIG. 29b is shown in more detail in FIG. 30 which develops this FIG. 29b, while FIGS. 31 and 32 illustrate (in more detail) portions of FIG. 30, namely respectively the unit B0 which does not has no input 20 RMAX and POSRMAX, and any of the units B1, B2 ... Bn

  which have an RMAX input and a POSRMAX input. Of course, it is possible to make memory block 600 only understand assemblies of the type illustrated in FIG. 32, using only part of the set of FIG. the input MUX multiplexer blocks).

  With reference firstly to FIGS. 29b, 30 and 31, it can be seen that B0 receives at input * the signal IN 107s, namely the output signal of the adder 107 with two inputs receiving, one, a validation signal. 30 VALIDATION and, the other, one of the output signals OUT of the blocks B0, B1, B2 .... Bn, C an ADRES-p address signal from the PLC 602, through the controller 601, * a calculation signal CALCUL, * an INIT start signal from an external sequencer (not shown), and

  the output of a multiplexer 105 serving as address signal ADRin for the blocks B0, B1, B2 ... Bn.

  s In the registers RMAX 704 and POSRMAX 705 of M0 (FIG. 31) are stored the values of the input signals IN and

(ADR-in) respectively.

  The input of a signal IN 107s, corresponding to the new accumulation, is validated by a comparator 706 with inputs P and Q, the first receiving the new signal IN and the second the contents of the register RMAX 704, if and only if P > Q, that is, if the value of the new IN signal is greater than the already stored RMAX, because in this case a new, higher RMAX value is entered. This registration of the new RMAX is performed through the AND gate 707 which produces, during the calculation cycle (signal CALCULATION), a WRITE registration signal applied to the input.

  recording In the register 704 of RMAX.

  At the same time, the address of the new RMAX is entered in the POSRMAX register 705 by its input En under the control of the WR registration signal from the gate 707 during the calculation cycle. The output signal OUT of the unit M0 is equal to "0" if the input En of the unit 708 comprising an AND gate is not valid, while it is equal to the contents of the register RMAX if this 25 entry is validated; the validation is carried out by means of a signal 709s constituting the output of the comparator 709 which compares the incoming signal ADR-in (a) and the content (b) of the POSRMAX register, this 709s validation signal being emitted by the comparator 709 if and only if a = b, that is, if the signal ADR-in actually represents the address or position of RMAX,

that is, POSRMAX.

  The block B0 of FIG. 31 also comprises an AND gate 710 with three inputs receiving the signal 706s from the comparator 706, the signal 709s output from the comparator 709, 35 after inversion, and the calculation signal CALCUL from the flip-flop 711 after the first CALCULATION operation after stopping the signal

RESET.

  Finally, in addition to its output signal OUT to the adder 107 (FIGS. 29b and 30), the unit M0 outputs, this time to the next block denoted M1: RMAXout, POSRMAXout, a signal ao, which is true if the signal IN is greater than the contents of the register RMAX 704, and finally a transmission signal TRo, which is equal to "0" for the processing of the first pixel of a sequence and which is true when the signal 706s leaving the comparator 706 is equal to 1.10, that is to say when a0 = 1, and when the ADRES signal is different from the contents of the POSRMAX register 705, that is to say when the

signal 709s represents a b.

  It is the INIT signal that triggers the RESET signal which,

  On the other hand, initializes, resetting them, the registers RMAX 704 and POSRMAX 705 and, on the other hand, the output of the flip-flop 711.

  As for the synchronism of the operation of the unit M0, this one is ensured by the front front amount of the pulses of a

  clock (not shown) defining the cycle of the pixel sequences).

  The four output signals RMAXout, POSRMAXo ,, t, a0 and TRo 20 of the unit M0 are inputted to the next unit B1.

  In FIG. 32, there is illustrated a representative unit Bi of a

units B1 to Bn.

  The unit Bi receives, on the one hand as the unit M0, the control signals IN, ADR-in, CALCUL and INIT (from which RESET is derived) and, on the other hand, RMAXin, (constituted by RMAXout of the previous unit),

  ai -., TR-1, and POSRMAXin, (consisting of the POSRMAXout of the previous unit), from the previous B unit (M1 receiving the output signals RMAXoUt, a., TRo and POSRMAXout of M0, while that the unit Mi receives the corresponding output signals from the unit Mi-1).

  The unit Bi firstly comprises two multiplexers 712j and 713i which make it possible to choose between two inputs for each of the registers RMAX 704i and POSRMAX 705j (analogues respectively

  to registers RMAX 704 and POSRMAX 705 of FIG. 31).

  Thus the multiplexer 712i chooses between the input IN and the value of the register RMAXIn of the unit B upstream (B0 or more generally Bi-1), while the multiplexer 713i chooses between the input ADR-in and the value of the POSRMAXin register contents of Unit B upstream. These two selections are determined by the selection signal SelMux for the multiplexers: the signals IN and ADR-in, received on the inputs 1 of the multiplexers, are validated by SelMux if the signal IN has a value greater than that contained in the register RMAX 704i (the comparison being made in the comparator 706i) and if the input signal ai1. from the upstream unit B is equal to "0", the AND gate 714i outputting a signal 714s if the comparator 706i outputs a signal and at the same time ai-1, before inversion, is zero. On the other hand, in the other cases (P is smaller than or equal to Q and / or A1-1 is different from 0), the multiplexers 7121 and 713 transmit RMAXin and POSRMAXIn of

  the unit Bi-1 (ie BO if BI = B1). It will be noted that the RMAX register 704i, the POSRMAX register 705i and the comparator 706i, on the one hand, the comparator 709i and the gate 707i; (whose roles will be specified below), on the other hand, are similar to the corresponding units of FIG. 31 without the index i.

  The OR gates 715i and ET 707i perform the validation of the write signal WR for the registers RMAX 712f and POSRMAX 713j when SelMux or TRI1. is valid and at the same time the signal

CALCULATION commands the writing.

  Finally, the AND gate 7101 (analogous to the AND gate 710 of FIG. 31) delivers a signal TR1 if simultaneously the output signal of the gate 715 is valid, the address signal ADR-in is not equal to the signal The output signal of the POSRMAX register 705e and the flip-flop 711j (similar to the flip-flop 711 of FIG. 31) is triggered by the first signal WR after the RESET signal resulting from the signal INIT has been stopped. The AND gate unit 7081 (similar to the unit 708 of FIG. 31) outputs a signal OUT consisting of the RMAX register content 704i to the adder 107 (FIGS. 29b and 30) when the

  comparator 709j found the equality a = b.

  Finally, the unit Bi outputs, in addition to the signal OUT, four signals in the Bi + 1 downstream unit, ie RMAXout, ai, TRi and PORSMAXout similar to the corresponding signals output by the unit Bo of FIG. 31. Further, as explained in FIG. 30, the signals RMAXout and POSRMAXout (simply noted RMAX, with an index of 1 to n, and POSRMAX, with an index of 1 to n) are sent on a block 617i (ie 6171 , 6172 ... 617n), these blocks, like block 617o, which receive the outputs RMAXout (denoted RMAXo) and POSRMAXout (denoted POSRMAXo) of B0, being activated by lo0 the address of index 0 to n of the demultiplexer 618 under the control of API 602. On the other hand, the last unit Bn delivers its RMAXn and POSRMAXn signals only to a block 617n activated by the address signal n of the demultiplexer 618 under the control of the PLC 602. It will be noted that is the inhibition of the signal RESET which authorizes the writing in the registers RMAX 704i and POSRMAX 705i at the end of cy INIT clock key. As for the signal TRi it is either "0" at the output at the first writing of said registers of the block Bi or "1" if the signal WR is valid and the signal ADR-in is different from the contents of the register POSRMAX 705i after the comparison done

in the comparator 709i.

  In FIG. 33, there is illustrated, in a schematic way, the

  implementation of STN modules, incorporating the class extraction method previously described in FIGS. 29 to 32.

  An STN module 660 of this type is controlled by a zone signal Z defining the analysis zone in the scene, and receives, during the sequence to, the motion signal MVT, this bi-linear type signal incorporates the information detection and speed of movement of the pixels of the scene. At the end of sequence to, during the END cycle, the different classes of the movement are extracted and entered in the memory M2, the demultiplexer 653 then debits the classes CMVT0, CMVT1, CMvT2 and CMVT3 during the following sequence t1. Each classification signal controls a multi-class STU type STN module, also referred to as a multi-class STN module, each receiving a bilinear x / y position signal. Thus, the CMVTO class signal controls the STN module 661, the CMVT2 signal, the STN module 662, the CMVT2 signal, the STN module 663, and so on considering the number of classes found. There are as many STN modules recruited as there are

classes found.

  At the end of sequence t1, the signals of class ZIMVTj of the STN modules receiving the signal x / y define all the zones associated by categories of movement. This implementation, in 10 two sequences, shows, in the first sequence, a processing of a signal in the time domain TD, then in the second sequence, a processing in the spatial domain SD from the previous classification results. from treatment

  previous in the time domain TD.

  During the sequences t2 and t3, the same principle is applied from each class ZjMVTj which outputs a control signal defining the analysis areas in the scene. In STNs of type 664, receiving a L / T / S tri-linear signal defining the color of the pixel and separating the different color classes of each of the incoming ZIMVTj classes, is

  performed TD time domain processing. The Downstream Type 665, 661, and 667 STNs separate into sub-zones each of the incoming classification signals from a bi-linear x / y position signal in a spatial domain SD process.

  This multiple class extraction process demonstrates the

  perception performance in speed of execution thanks to the implementation of multi-class extraction means. After four sequences t0 through t3, the amount of extracted areas is far superior to previously defined methods and devices.

  After the description of the sets and subsets

  FIGS. 17, 18, 19, 20, 22, 23a-23b, 23c and 25 with representation of the results in FIGS. 21a, 21b and 24, which illustrate the obtaining of the necessary information concerning the nature of the OB object observed and the position of the latter in the context of the use of a single representative signal, namely the luminance L, will be maintained with reference to FIGS. 34 and following improvements to make the determination of the relative location of two objects, the implementation of trees 5 representing the relative positions of the barycentres of the different areas, the implementation of three color components and finally the storage and recognition of perceived objects with invariance in translation, size and rotation . In FIGS. 170 to 62, the average o-positions, also referred to as BarZ1 barycentres, are essentially implemented as parameters or dimensions applied to the inputs of the STN (1), STN (2) and

  STN (3) constituting unit 8 of Figures 17 and 18.

  FIG. 34 schematically illustrates the use of an STN module (2), of the type illustrated in FIGS. 23a-23b (optionally with the modifications of FIGS. 23c and / or 25). ), and two orientation units poc and pj these two orientation units receiving the coordinates x and y and an orientation angle or rotation, ie ca for the orientation unit 150 and P for the orientation unit 151, so as to rotate the coordinate axes, from the initial position determined by the x and y axes to a derived position defined by the angles a and 13, or respectively the slopes pca and pp, these orientations and

  slopes being shown in Figure 34a discussed below.

  The STN module (2) of FIG. 34 thus receives, as input data, not DATA (Ax) and DATA (Ay), but rather DATA (Apa) and DATA (App). The processing carried out in the STN module (2) is controlled by a program register 152. Finally, the signals collected on the bus 114 (taken from FIG. 23a-23b) ultimately represent not only the Zs, but especially the 30 centroids, ie BarZi depending on slopes pa and P3,

  i.e. angles ca and p respectively.

  In FIG. 34a, two particular values of the angle α 1 are illustrated, namely (x and cX 2, and two lines of slopes pal and pOE 2 relative to the direction of the x axis, representing the processing in the unit shown in Figure 34. The slopes pcc9 and poa10 are respectively perpendicular to poal and poa2 slopes so as to achieve cartesian coordinates of slopes paol

  and pa9 for the angle a1 and pa2 and pal0 for the angle ao2.

  In Figure 34a, there is also illustrated the terminals a and b, 5 corresponding to the slope pa2, and c and d, relative to the slope pox9. The object OB is thus defined by its barycenter BarZ0 (the index "0" indicates that it is the first center of gravity or barycenter of origin of the tree discussed below) in the rhombus referenced 160 (constituting the initial zone Zo) defined by the lines of

  o coordinates a, b, c and d delimiting this object.

  By varying a, therefore the slope pa, a rotation is made of the straight lines A and B as illustrated in FIG. 35, in which there is shown, in the frame or zone Z0, different orientations of the angle α, which which corresponds to Zr variables namely ZrO (for a = 0), ZrO10 (for a = 10), Zr30 (for oa = 30) and Zr170 (for oa = 170). In this figure 35, there is illustrated in particular the slope corresponding to the angle of 30, namely the slope pa30 with corresponding terminals A and B corresponding to a and b for the value of the angle a = 30. It can be seen from Fig. 35 that, as expected, all BarZr bands pass through the BarZ0 center of gravity as defined in Fig. 34a. In Fig. 36, which illustrates a variant of a portion of Fig. 23a, namely unit 101 thereof, there are comparators 110a and 111a, on the one hand, and 110b and 111b. b, on the other hand, as well as AND gate 112a (which receives as input the direct outputs of comparator 110a and inverted comparator 111a) and AND gate 112b (which receives the direct outputs of comparator 110 Ob and inverted comparator 111b ). There is also in FIG. 36 the AND gate 131 receiving the outputs of the AND gates 112a and 112b. The novelty consists, in the case of FIG. 36, with respect to FIG. 23a, in providing, in addition to the AND gate 131, an OR gate 132 which receives, as the AND gate 131, the outputs of the AND gates 112a and 112b. The outputs of the AND gates 131 and OR 132 are sent to a multiplexer 133 controlled by a selection signal 134 (controlled by the STN module PLC) between the output of the AND gate 131 and that of the OR gate 132 and finally gets at the output of the multiplexer 133 a signal 101s applied

  (consisting of 131s or 132s) to bus 114.

  Finally, Fig. 36, like the corresponding part of Fig. 23a, relates to a classification subassembly (101 in Fig. 23a and 101 'in Fig. 36) in bi-linear mode, i.e., for a STN module (2) receiving two input signals namely DATA (Ay) and DATA (Ax). In the case of FIG. 23a, the 10 four comparators 11a, 11b, 111a and 111b make it possible to output four signals, namely ax, bx, ay, and by respectively. Due to the logic functions performed by the three gates 112a, 112b and 131, the resulting signal 101s delivered by the AND gate 131 is defined by

101s = (ax.bx). (Ay.by).

  Since the Boolean operation above is commutative, we can transform it into

101s = (ay.by). (Ax.bx).

  Referring now to FIG. 36, which illustrates a variant of FIG. 23a, it can be seen that the assembly illustrated in this FIG. 36, carries out, at the output of the door 131, the same Boolean operation AND that the whole of the FIG. 23a, the output signal 131s being equal to that indicated above. However, considering the OR gate 132, which includes the whole of FIG. 36, a wider range of application of the output signal of the gate 132 relative to that of the gate 131 is obtained. , the output signal 132s extends

also at (ay.by) + (ax.bx).

  This difference is illustrated by comparing the domains of FIG. 37, which corresponds to the assembly according to FIG. 23a, and FIG. 38, which corresponds to an assembly according to FIG. 36. In these two FIGS. 37 and 38, FIGS. the x and y coordinate axes, as well as the Ax and Bx, Ay and By terminals. In the case of FIG. 37 (assembly according to FIG. 23a), the output signal of the AND gate 131 (signal 101s) which arrives directly in the bus 114 has for its domain the grayed-out rectangle Zet delimited by the four aforementioned terminals because the AND gate 131 passes only the signal included between both Ay and Bx terminal and between the Ay and Borne By terminals; on the other hand, in the case of the circuit according to FIG. 36, the signal 101s is constituted either by the signal 131s or by the signal 132s; in the first case, the domain of 101s, namely that of 131 is that illustrated in Figure 37; in the second case, the domain 101s, namely that of 132s, is the one shown in shaded form ZOU in FIG. 38, this domain being comprised either between the terminal Ax 10 and the terminal Bx or between the terminal Ay and the terminal Borne By, simply eliminating the rectangles left blank in FIG. 37, in which the signal is simultaneously less than Ax, as regards its x-coordinate, at the Ay-bound, with respect to its y-coordinate, or greater than both at the B-terminal , with respect to its x coordinate and to Borne By with respect to the y coordinate. The additional OR gate 132 of FIG. 36 thus increases the range of application by

  compared to that of Figure 23a not including such a door.

  The multiplexer 133 of FIG. 36 makes it possible to choose between the two solutions, namely between the output 131s of the AND gate 131 (greyed area in FIG. 37) and the output 132s of the OR gate 132 (domain greyed out on 38) according to the case that one wishes to treat and this under the control of the signal 134 of

  control of this multiplexer (from the API).

  In the case illustrated in FIG. 35, which corresponds to the illustration of the only AND gate 131, numerous narrow band or "line" orientations delimited by terminal A and terminal B were used for different orientations or slopes, for example. example po30 for the band Zr30, the center of gravity BarZ0 of the zone Zr being defined by the point of intersection of the bands ZrO (slope 0) to Zr

(slope 170).

  Instead of using a large number of narrow strips, the Z0 plane can advantageously be covered by means of sectors, as illustrated in FIGS. 43a, 43b and 43c, which correspond to the use of the only OR gate 132 of FIG. FIG. 36. After a first sequence of determination of a first Z0 and its barycenter BarZ0, the device starts a second sequence according to FIG. 39, in which the grayed portion corresponds to the shaded portion of FIG. the axes instead of being orthogonal, form an angle (acute) defined by cal and a2, while the terminals, instead of being defined by Borne Ax, Borne Bx and Borne Ay, Borne By are representative of the position of the BarZ0 center of gravity (included in Z0) by the slopes pal and

  pa2. In this FIG. 39a, there is shown the terminals Terminal Apal and Terminal Bpal for the slope pcal and the terminals Apca2 and Bpa2 for the slope pa2, the zone Zrl being that seen by the module STN (2) of FIG. signal 101s derived from the output signal 132s of the OR gate 132.

  In the third sequence illustrated in Figure 39b appears an additional signal representative of BarZ1. As a result, the computation cycle of the STN module (2) of Fig. 34 determines the position of BarZ1 in addition to BarZo, while the update cycle of the results determines the new terminals Terminal A'pcal, and B pcal, on the one hand, and A'pco2 and B'pcc2, on the other hand, defining BarZ1. In a fourth sequence (FIG. 39c), the sectorization is refined by dividing the sectoral zone Z21 determined in the course of the third sequence into several sub-zones.

  sectoral Zra, Zrb, Zrc Zrd by the addition of an STN module (2) per sub-area operating in the same manner as the STN module (2) implemented according to Figure 39a; as a result, the slopes pal and poc3 are replaced by intermediate slopes limiting the aforementioned sub-zones.

  At the end of this fourth sequence, one of the STN modules (2)

  sub-area recovered BarZ1, namely the Zrc subfield (Figure 39c); the bisector of this particular subzone Zrc, defined by two very close slopes, accurately determines the poa3 slope illustrated in Figure 39c.

  Considering the slope pOc4, perpendicular to pac3, we see that the bounds A and B bind BarZ0 of the zone Z0 at the intersection with the band perpendicular to pa3 bounded by bound A and bounded B, while the limits bounded A 'and' B 'terminal allow to define BarZ1 at the crossing with the band perpendicular to poa3 limited by Terminal A and Terminal B (this band including both BarZ1 and BarZo). The distance between the axes of two bands perpendicular to pca4 passing through BarZ0 and BarZ1 and perpendicular to the BarZo-BarZ1 band represents the distance p3 between these two barycentres (that corresponding to the angle oE3);

  The angle c3 of the slope pa3 and the distance p3 are the two polar coordinates of this centroid, such as BarZ1 with respect to the centroid BarZ0 taken as the origin of the coordinates, and an axis (not shown), for example parallel to the lower edge of zone Z0 (FIGS. 39a and 39b).

  For subsequent sequences, all STN (2) modules are

  released except the STN module (2) having recovered the Zrc subfield.

  This block recorded the results p3, ac3, Cox3 in registers 1070, 1071, and 1072 respectively. These registers belong to the results register area 104 of the STN module.

  Between BarZ0, which is the center of the Z0 zone, and BarZ1, which

  is the centroid of the zone Z1 included in the zone Z0, we have the relation 49 defined in FIG. 40, BarZ0 representing the "father" from which the "son" BarZ1 is derived, c3 and p3 being the polar coordinates of the son by report to the father.

  This four-part description shows the method of

  determination of the pux link between two centroids from a dynamic recruitment of STN (2) modules during sequences two to four. The STN modules (2) other than the STN module (2) selected in the fourth sequence, namely that corresponding to the Zrc subfield, become available. The same is true of STN (2) modules not selected in the

previous sequences.

  In FIG. 41, there is illustrated a unit M (0) consisting of a pair of STN modules, the first module 296 being a monolinear STN module (1) with a DATA input parameter (A), simply denoted A, at 296a, which is processed by a desired function FoG 'to feed a group of analysis output registers, noted reg 296b, in which accumulates representing, in the form of a histogram, the statistical distribution of the DATA parameter ( AT); the output signal, at 296c, of the STN module (1) 296 is a classification signal CA 350. The second module 297 is a bilinear module STN (2) with two input parameters, namely the x and y coordinates applied at 297a, which are processed by a desired function FoG "of analysis output, feeds a group of registers of analysis output noted reg 297b classifying in histograms the two parameters x and y; an output signal of the module STN ( 2) 297 consists of the zone Z0 and is fed back as a signal 297s to an auxiliary input 296d of the first module 296, while another output signal of the register 297b is constituted by BarZ0. the modules STN (1) 296 and STN (2) 297 are advantageously constituted as illustrated respectively in FIG. 20 and in FIG. 23a-23b (optionally with the variants of FIGS. 23c, 25 and / or 36), in order to finally get as explained with reference to these

  figures, Zo output signals and BarZo.

  In FIG. 42, which represents a set of STN modules capable of executing the successive operations described above with reference to FIGS. 39a, 39b and 39c, there is the unit M (0) of FIG. pair of blocks STN (1) 296 and STN (2) 297, the input signals denoted A, x and y and the output signals Z0 and BarZ0, to this first unit M (0) of two STN modules (1) 296, and STN (2) 297 determining Z0 and its barycenter BarZo are associated, on the one hand, in succession of analogous units, of which only the first one has been illustrated, namely M (1) which determines Z1 and BarZ1 from Z0 and parameters B (and no longer A), x and y (coordinates); The following unit (not shown) at two STNs determines Z2 and BarZ2 from Z1 and parameters C, x and y; subsequent units also not shown determine Z2 and BarZ2, etc .; and on the other hand, a set M (2) of STN modules (2), representing the dynamic recruitment, denoted 300, 301, 302, 303 ... 307 respectively receiving a pair of slopes p0 and pl, pl respectively. and p2, p2 and p3, p3 and p4, ... p6 and p7 to respectively determine ZrO, Zr1, Zr2, Zr3 ... Zr7 (illustrated in Figure 43a) from Z0 and BarZO from the unit M (0), and BarZ1, from the unit M (1), whose three values are received as input by each of the units 300 to 307; The set M (2) is accompanied by similar unrepresented assemblies each receiving ZO and BarZO from the set M (2) and furthermore BarZ2, for the set of M (2) type of rank just after the M together (2) illustrated, BarZ3 for the unit M (2) of rank just after and so on, each of the STN (2) of each of these sets M (2) not shown receive a pair of slopes to the manner of the STN (2) 300 to 307 illustrated: these successive units of type M (2) determine at the output of their modules similar to the modules 300 to 307 illustrated signals defining sectors of type ZrO to

  Zr7, but relative to Z1, Z2 etc ... and no longer to Zo.

  The operation of the assembly illustrated in Figure 42 is as follows.

  During the first sequence (video image or frame of the video image for example), which involves the set M (0), the classification signal 350, developed in the module 296, arrives in the module 297 operating in two-dimensional mode with the x and y coordinates (defining the pixel of the video image) as input parameters (in addition to the signal 350). The result of the calculation of the histogram in the module 297 controls the function FoG "of automatic classification with anticipation upon the arrival of the signal 350, that is to say during the END phase of the end of the sequence. calculation and classification in the register reg 297b of the module

  297, allows (as explained with reference to FIG. 34) to determine, on the one hand, Z0 by means of the update of Borne Ax, Borne Bx, Borne Ay and Borne By flanking Z0 and, on the other hand, BarZO using POSMOYX and POSMOYy. The first sequence 35 thus makes it possible to determine Z0 and BarZo.

  During the consecutive sequence, namely the second sequence, the set M (2) recruited for the second and fourth sequences comes into play, the STN modules (2) 300 to 307 (of which only some modules have been illustrated), programmed to receive as input a pair of steered slopes of 22o30 'at 22o30' between 0 (slope p0) and 157 30 '(slope p7) each slope of index j representing an angle of 22.5j - perceive BarZo at inside the zone Zo according to their degree of orientation (Figure 43a). At the end of this second sequence, at the time of the new end-of-sequence END signal, the histogram calculations in STN (2) modules 300 to 307 make it possible to update the classification terminals. B and therefore define oriented areas called search zones ZrO to Zr7, as illustrated in Figure 43a for zones Zrl, Zr2, Zr3 ... Zr7 15 determined by STN blocks (2) 300, 301, 302, 303 ... 307, respectively. FIG. 43a actually represents the accumulation of the oriented zones Zr defined by the modules 300 to 307 during the second sequence. The second sequence allows to position the sectorization of the plan around BarZo. In addition, at the end of this second sequence a related feature appears

  to the first classification signal 350, namely a sub-area characteristic (not shown) generating a BarZ1 center of gravity (illustrated in FIGS. 43b and 43c which correspond to FIGS. 39b and 39c) which is used to determine the module p and the orientation angle 25.

  In the third sequence, which immediately follows the second sequence, the unit of STN (2) modules 300 to 307 receives, in addition to Z0 and BarZo, a third signal BarZ1 while continuing to receive the slopes po, Pi. .. p7; pi equivalent to the difference between the two classification terminals, namely from these inputs, one of the STN modules of M (2) - for example 303 - reference (FIG. 43b) an additional centroid BarZ1 and becomes at the end of the sequence , the STN module selected to operate in the

  next sequence, namely the fourth sequence.

  During this fourth sequence, the inputs of the STN modules (2) 300 to 307 reorganized as represented in FIG. 39c while receiving: * the slopes p3, pi ..... pj, p4 representing several sectoral sub-zones of the zone Zr3 , and * the slopes p'3, p'i ..... p'j, p'4 respectively perpendicular to the previous slopes, * p and a are calculated by the selected module - for example 303 -,

  p being the distance BorneA'-BorneA (Figures 39c and 43c).

  Then a new cycle of four sequences can start over the sets of module STN, analogous to the set M (2), but not represented to determine Z20 and

  BarZ20 and so on, using the released STN modules.

  In the case of an object recognition, a pca portion of BarZ1 can be proposed and the sequence four is applied directly after the sequence two, the sector Zr3 is subdivided, the module p and its angle cx are calculated with a sequence beforehand. Reference will now be made to FIGS. 44a to 44e, 45 and 46, on the one hand, and 47a to 47d, 48 and 49, on the other hand, to examine two particular cases of relative disposition of zones Z10o, Z20 AND Z30. inside the zone Z0, namely respectively a nesting in the manner of the Russian dolls (but in two dimensions), on the one hand, and separate locations, on the other hand. In this set of figures, FIGS. 44a to 44c and 47a to 47d represent the sets M of STN modules, FIGS. 45 and 48 the corresponding relative provisions of the zones, and FIGS. 46 and 49 show the BarZo to BarZ3o baritone centers of these boxes. areas. We will first examine the case of nested areas according to Figure 45 with implementation of M sets of STN modules for the two parameters x and y according to Figures 44a to 44e. We will begin by studying the creation of a link (according to Figure 46) between the BarZ0 and BarZ10 centroids of the two nested zones Zo 35 and Zo10 (according to Figure 45). First, during the phase to (Figure 44a). ) the parameter Data (A), denoted A, determines, thanks to the module M (0), a

Zone Zo and its center of gravity BarZ0.

  In a second step, during the phase t1, the signal Z0 produced by M (0) in the phase Co acts on the one hand on a module M (1) locating the subarea Zo10 triggered by a data datum (B), denoted B, in order to determine Z10 and its centroid BarZ10 and, on the other hand, by an M module (2) operating as explained above with reference to FIGS. 42, 43a, 43b and 44c and 1o grouping the different orientations Po, Pl ... Pm, in order to

  position these orientations around BarZo center of gravity.

  In a third step, during phase t2, the module M (2) receives BarZ10 (in addition to BarZo) which has just been determined by

  M (1) and search for the sector that contains BarZo10.

  In a fourth step, during the phase t3, the aforementioned sector having been found, an STN module (2) denoted SBi (FIG. 44e) among the STN (2) modules of the set M (2) is selected, this which determines the value of i corresponding to the angles of the

  bisector of the subsector selected by BarZlo.

  The dynamic grouping of M (2) disappears in favor of the SBN bilinear STN (2) module selected. Registers 1070, 1071

  and 1072 respectively containing p, oc and Cax are updated.

  In FIGS. 44b to 44e, we have noted la, Ila, Illa and IVa.

  structures involved respectively at times to, t1, t2 and t3.

  The operation that has just been described, to determine Zo10 and BarZ10, relative to Zo and BarZo, can then be applied for each of the other nested zones, that is to say to determine Z20 and BarZ20 from Z10 and BarZlo , then Z30 and BarZ30

  from Z20 and BarZ20 so on.

  The structures involved at times t1, t1, t2 and t3 of the subsequent phases make it possible to determine the zones Z20, Z30 and their centroids. In Figs. 44c, 44d and 44e, lb, lib, llIb, IVb,

  then Ic, lic, lllc, followed by IVc (not shown) and finally Id, lid followed by Ilild and IVd (not shown).

  We finally obtain the linear tree (without branches) of Figure 46 illustrating the branching of the BarZo -> BarZ10 -> BarZ20 -> BarZ3o barycenters, with the polar coordinates p, oc of a downstream barycentre (son) with respect to a upstream barycentre (father); as an example BarZ2o is the son of BarZ1o. The shaft of FIG. 46 illustrates the arrangement with successive interlocking of the corresponding zones Z0, Zo10, Z20, Z30 of FIG. 45, Z30 being housed in Z20 which is housed in Zo10, itself housed inside Z0. As an example in the case of a human figure, we examine successively the whole face (Zo), then only one eye (Zi), then the

  pupil (Z20) and finally the iris (Z30). The tree is representative of the deepening of the characterization of the examined object, independently of the rotation thereof, because the successive polar coordinates of the father-son pairs have been determined.

  After examining the case of a nesting of the successive zones, therefore a linear tree without branching, with reference to FIGS. 44a to 44e, 45 and 46, we will study the case where the zones Z10, Z20 and Z30 are located at Z0, but are separated from each other (Fig. 48), with reference to

Figures 47a to 47d, 48 and 49.

  If Figs. 47a and 47b are identical to Figs. 44a and 44b, respectively, so that Z0 and then BarZ0, Zo10 and BarZo10 are also obtained at first and second times respectively, since Zo10 is contained in Z0. , as the case of FIG. 45, on the other hand, the continuation of the treatment is different, since the zone Z20 in this second example of FIG. 48 is not contained in the zone Zo10, but is distinct.

  of it, however, being contained in zone Z0.

  In FIG. 47c, at the third time t2, the module M (la)

  analogous to module M (1), but with C at the input instead of B, also receives Zo10 as in the case of FIG. 44c (since now Z20 is included in Z0, but not in Z1o) and determines Z20 and BarZ20, while the module STN M (2a) determines the p corresponding to a zone Z20-

  In a fourth time t3 (Figure 47d) and the following times, the process is repeated for new zones arranged inside the zone Zo, but separated from each other, using, either the Z0 of the previous phase as in the case of a nesting of the successive zones, but Zo as

  that input modules such as M (9).

  We thus successively determine Zo, Zo10, Z20 and Z30 and we

  gets a tree with branches or branches from BarZ0.

  As an example, we examine the whole face (Zo), then the mouth 10 (Zo10), then the right eye (Z20) and finally the left eye (Z30), the label face encompassing the labels mouth, eye right and left eye. The letters A, B, C correspond to the various curves a ', b', c 'of FIG. 15. After having examined the case of successive appearances, only one after another, of zones, are nested in one another. other, or separated, in correspondence with a tree without subembranchements, we will consider the most general situation in which two or more zones appear simultaneously during the same sequence (constituted by one or more frames or images of a video signal), in correspondence with a tree with sub-branches or branches. For the case of several peaks or classes appearing in the same sequence, we can now deal with the case where the Z-type zones are some nested and some separated, that is to say the intermediate case (and more general) between the boundary cases of the nested zones (figures 44a to 44e, 45 and 46) and the separated zones, but contained in the initial zone Z0

(Figures 47a to 47d, 48 and 49).

  Figure 50 illustrates such an intermediate case, in which there is

  at the same time interlocking of at least one zone in another, for example of Z21 in Z12 (Z12 -> Z21), and positions separated from certain zones, such as Z11, Z12 as well as Z22 and Z21, all these zones being contained in zone Z0, as is also seen on the shaft of FIG. 50a.

  FIG. 50 shows the first processing steps during refocusing, or more generally increasing the resolution, with the abscissa, from left to right, the durations of several sequences or groups of successive sequences to, t1, t2 and t3, and on the ordinate, from top to bottom, the succession of treatments performed on each

sequence or groups of sequences.

  1) Considering first the ordinates, that is to say the successive rows of Figure 50: _0 The first row of Figure 50 illustrates, from left to

  right, the successive appearance of zones due to the increase of the resolution according to the parameters: first Dif (difference between the initial sequence with maximum resolution r max during To and the subsequent sequence at minimum resolution r min 15 during T2 or to (Figure 3), then a, b .... (Figures 14 and 15).

  The second row of Figure 50 illustrates the first three successive treatments performed on the first perceived area,

know Zo.

The third row of FIG. 50 illustrates the first two successive treatments performed on the two sub-areas.

  determined after the first zone Z0, namely the two subzones Z11 and Z12.

  The fourth row of Figure 50 illustrates the first of the

  successive treatments carried out on the two sub-zones 25 determined next, namely Z21 and Z22, Z21 being included in Z12.

  2) Now considering the abscissae, that is to say the successive columns of FIG. 50, namely the sequences or groups of successive sequences: The first column, corresponding to the initialization at time 30 to, represents the appearance of the first zone Zo following the first increase of the resolution (after its abrupt decrease), during the phase Ti of figure 3), and the determination of the

BarZ0 center of gravity of this zone Z0.

  The second column represents the two simultaneous treatments on the following sequence at the time t1 and illustrates the continuation of the treatments of the zone Zo, namely: on the one hand (in its first row), the appearance of the sub-zones Z11 and Z12 within Zone ZO, with their respective centroids BarZi1 and BarZ12, and - secondly (in its second row), the segmentation of zone ZO into sectors, as explained above with reference to

Figures 42 and 43a, 43b, 43c.

  The third column represents the three simultaneous treatments performed on the same sequence, that at time t2, namely: - the appearance (in its first row) of the two subzones Z21 and Z22 resulting from the increase in the resolution, with their respective barycenter BarZ21 and BarZ22, - the determination (on its second row), in each of the sectors of ZO (those of the same row, but of the preceding column), of the barycenter BarZ11 or BarZ12, and of the module p and of the associated angle ax constituting the polar coordinates of BarZ11 or Bar12 with respect to BarZo, the BarZo and BarZ12 vectors rays being referenced 50 and 51 respectively, and - the segmentation (in its third row) in sectors of

  zones Z11 and Z12 from the respective barycenter BarZ1 and BarZ12.

  The fourth column shows the continuation of the same type of treatment, namely: in the first row, the appearance of subzone Z21 (included in Z12) and subfield Z22, with their respective barycenter BarZ2, and BarZ22 in the second and third rows, the determination of the modules p and the corresponding angle α for the positions of BarZ22 and BarZ21 with respect to BarZo and BarZ12 respectively, the vector radii being denoted 52 and 53 respectively, and - in the fourth row, sector segmentation of

zones Z21 and Z22.

  FIG. 50a shows the tree constructed from the processes carried out in FIG. 50. In this FIG. 50: the first parameter Dif made it possible to determine the zone Zo and its centroid BarZo, which serves as a point of origin to the (inverted) tree, - the second parameter allowed the appearance of the two subzones Z11 and Z12 (included in the Zo zone), with their respective barycenter BarZ11 and BarZ12, the father / son relations between the father BarZO and the wires BarZ11 and BarZ12 being materialized by the vector rays 50 and 51 respectively, with the angles a1 and Oz12 and the modules p11 and P12, and the third parameter ba made it possible to show two new sub-zones Z21 and Z22, the first included in Z12 (and hence in ZO) and the second in ZO only, with their respective BarZ21 and BarZ22 barycenter and the father-son relations 53 and 52 from the BarZ12 and BarZO fathers respectively, as well as

  the a and p corresponding to 0 (21, 0.22 and P21 and p22.

  FIG. 51 partially illustrates an application of the multiple zone selection developed in FIGS. 42 to 50, from spatio-temporal parameters obtained during the refocusing, or more generally from the restoration of the resolution from its minimum value to its maximum value, as indicated above with reference to FIGS. I to 16; the

  FIG. 51 corresponds to the periods T2 and T3 of FIG.

  During the period T2, namely the sequence t0, a signal is used, in particular of difference Dif corresponding to w = 20. This difference signal is introduced on the input Dif of the first one-dimensional module STN (1) 90 of the module M (0) which also comprises a second two-dimensional module STN (2) 91. At the end of the period T2 = t0, the reference zone Z0 has been identified, as well as its center of gravity BarZo, by the module STN (2) 91 of the set M (0). The period T3 refocusing, or in general passage from the resolution of its minimum value to its maximum value, comprises sequences t1, t2, t3 ..., only the sequences t1

  (corresponding to w = 16) and t2 (corresponding to w = 12) being illustrated in Figure 51. During these periods t1, t2 ...

  appear in turn the series of spatio-temporal signals a, b,

c ... of figure 15.

  It can be seen that Figure 51 shows Figure 47d with the exception of dynamic recruitment modules to calculate the p and associated for clarity. During the first sequence T1 of the period T3, the signal a w = 16 appears and is analyzed by the set M (1) comprising the two modules STN (1) 92 and STN (2) 93; at the end of this sequence tl, a multi-class extraction analysis and a descending order distribution, as previously explained (with reference to FIGS. 31 and 32), of each of the classes found by a two-dimensional module STN ( 2) associated, namely the second class on the module STN (2) 94, the third class on the module STN (2) 95 and so on for the modules

STN (2) 96, STN (2) 97 et seq.

  During the second sequence t2 of the period T3, which sequence immediately follows the sequence t1, it is the signal b (w = 16) which appears and which is analyzed by the set M (2) with two STN modules (1 192 and STN (2) 193; during this sequence t2, a descending distribution of the classes found on an associated two-dimensional module STN is also carried out, in this case: the second class on the STN module (2) 194, the third 25 on a bidirectional module (no shown) downstream of the STN module (2) 194, as the STN module (2) 95 is disposed downstream of the STN module (2) 94, and so on other modules (not shown but illustrated by the vertical arrow) located under the

STN module (2) 194).

  During the subsequent sequences t3, t4, ... of the period T3, analogous treatments take place, at increasing resolutions w = 8, w = 4, w = 0, in blocks similar to the blocks M (0), M (1) and M (2) and in two-dimensional modules STN (2) similar to the modules 94, 95, 96, 97 shown in Fig. 51, which stops at the resolution w = 12. Zones Zi and their corresponding BarZi barycenter are used for subsequent treatments, as previously described, FIG. 51 illustrating only zones Z0, Z11, Z12, Z13, Z14, Z15, Z21 and Z22, with their associated barycenter BarZo,

  BarZ1, BarZ12, BarZ13, BarZ14, BarZ15, BarZ21 and BarZ22.

  At the top of FIG. 51, the images and the zones corresponding to the three sets M (0), M (1) and M (2) are illustrated for the resolutions defined by w = 20, w = 16 and w = 12; several areas are illustrated on the images for resolutions defined by w = 16 and w = 12; in FIG. 51, only the first three columns have been illustrated, the arrow facing the image corresponding to w = 12, schematizing the fact that there are columns to

  right of the last column shown in Figure 51.

  Fig. 52 illustrates an improvement in the arrangement of Fig. 51 to lock the information to overcome the fact that the parameters a, b, c ... are temporary. For this purpose, in the embodiment of FIG. 52 (on which we find the STN (2) 93 to STN (2) 97 modules of FIG. 51 processing the x, y coordinates and the STN (1) module 92, with their inputs and outputs), each of the sub-areas is locked; the modification consists in adding, in the embodiment of the

  FIG. 52, additional three-dimensional STN (3) modules processing the color components L, T, S and referenced 93a, 94a, 95a, 96a and 97a whose input is connected to the classification output D of the STN modules (2) referenced 93, 94, 95, 96 and 97 corresponding.

  Locking on the dominant color is performed by analyzing the dominant color of each of the subfields by the STN (3) 93a to STN (3) 97a module having as inputs L, T, S, i.e. three color components, and as Zi zone control for each of these subzones; at the end of a sequence, the histogram of the components L, T, S makes it possible to define classification boundaries in L, T, S and thus to output a classification signal 101s (FIG. 23a-23b, or possibly 23c, 25) corresponding to the dominant color; the lock is obtained by connecting this classification signal, corresponding to the dominant color, to the input of the corresponding STN module which

calculates the associated area.

  In FIG. 52, the pairs of modules 93-93a, 94-94a, 95-95a, 96-96a and 97-97a are represented symbolically on the right with position (x, y coordinate) and color (FIG. components L, T, S) and outputs Ci, C12, C13, C14 and C15 representing the dominant color, from the parameters calculated according to Figure 51, namely Z11 and BarZ11, Z12 and BarZ12, Z13

  and BarZ13, Z14 and BarZ14, Z15 and Bar15.

  Finally, in the last column of FIG. 52, the elements allowing a representation by means of a tree are represented by illustrating the tree nodes BarZij (ij successively representing 11, 12, 13, 14, 15), a node being characterized by

  its zone Zij and its dominant color (Cij) associated.

  So far, it has been explained how, in the context of the invention, it was possible to perceive objects. Referring first to FIG. 53, we will now explain how, again in the context of the invention, it is possible to characterize the shape of the localized object, in particular in order to be able to control whether a new form is analogous to a form already determined, the characterization of the perceived object with regard to its form being carried out in two stages: it is first of all a question of coding the information on the object concerning its characteristics, then to check if a new object corresponds to the characteristics of a previous object, already perceived and coded or if it is new. It is first recalled that thanks to the assembly illustrated in Figures 44a to 44e, 45 and 46, on the one hand, and the assembly shown in Figures 47a to 47d, 48, 49, other On the other hand, it was possible to acquire different angle pairs a, (or slope p), and modulus p corresponding to a characteristic in zones Z0 and following. In FIG. 53, the bilinear modules m (0), m (1), m (2) ... m (j) ... homologous to the bilinear module SBi of FIG. 44a are illustrated in the left column. which expresses the result of dynamic recruitment in four stages according to this figure; the inputs and outputs of the modules m (0) ... are the same as those of the module Sbi; in particular the outputs of these modules m (0) ... deliver all the aforementioned couples a, p perceived in the zone Z0 (with an index 0, 1, 2 ... j ...) which are applied to a double bus BB which serves as input to a set MM which constitutes the essential means for determining a representation of the

  characteristics of the perceived object.

  The set MM consists of several subsets 10 MM (0), MM (1), MM (2)... MM (i) ..., each of these sets comprising two STN modules, the first of type two-dimensional referenced 801, 802, ... 80i ..., which receives as input, from the double bus BB oa, p, and the second of the one-dimensional type referenced 810, 811, 812, ... 81i ..., which receives a LABEL signal whose role will be explained below. In each set MM (0),

  MM (1) ... with two STN modules, the first STN module (2) 800, 801, 802 ... 80i ... debited in the second STN (2) 810, 811, 812 ... 81i .. .

  whose output consists of the indication LABEL0, LABEL1, LABEL2 ... LABELi ..., which characterizes the shape of the perceived object and which, on the one hand, is reinjected to the retro-annotation input of the associated first module STN (2) in the same subunit MM (1). ), MM (2) ... MM (i) of the set and, on the other hand, is the API which calculates, in particular (the relatively long time available for this calculation) by a RMAX terminal software and POSRMAX, the histogram of labels LABEL0, LABEL1 ... received

  successively during the whole period T3, the increase of the resolution and which determines, at the end of this period T3, the position of the maximum quantity of identical labels corresponding to the characteristic label of the perceived object, noted LABEL GLOBAL at the bottom 30 of Figure 53.

  Finally, the indication of the order (0, 1, 2 ... i), framed by a square, defined by the first module STN (2) of each subassembly constitutes the input of the second module STN (1) of all

  subassemblies by the LABEL L bus.

  It will be noted that all the operations of the assembly of FIG. 53 concern the signal of the zone Z0 which is applied to a second feedback input of the different first STN modules (2) of the sets MM, ie 800, 801, 802 ... 80i 5 ..., while the API manages, by means of a software, the different

  sequences 0, 1, 2 ... i ... related to functions.

  After initial zeroing of all the memories of the STNs of the set MM, as well as the value of the pointer of the last acquired object, the device of FIG. 53 can operate. A sequence number, from 0 to n (passing through i) by incrementing by 1, is assigned to each of the sets MM (0), MM (1), MM (2) ... MM (i). up to MM (n) (not shown). It should be noted that for the first subset MM (0), corresponding to the zero order number, the classification terminals of the two STN modules 800 and 810 are open to the maximum, which

  applies "0" to the maximum value of the input bus.

  During the process, the resolution changes from its minimum value to its maximum value: * during the period T2, the calculation of the bilinear histogram, namely the memorization of the characteristics of the object, is blocked with reset to zero. the unilinear histogram of each of the subsets MM (0), MM (1), MM (2) ... MM (i) ...; during the period T3, the operating cycle begins with, for each appearance of a new pair o, p the test of whether the pair is known or not; two cases are possible: - in the case where the pair a, p does not find, in the subsets MM (1) to MM (n), any classifier which is valid except that defined by the pointer indicating a "last acquisition", it is the learning phase which is activated and which consists in the fact that the subset MM of last acquisition, which has its classifiers still valid, forces the validation of its associated LABEL on the bus L of LABEL and the histogram of the first module (800 to 80i) of the last-acquisition MM subassembly 820-82i stores the associated ax, p, value 820 to 82i pair. Such an operation is repeated for any new pair a, p during the period T3 of transition from the minimum resolution to the maximum resolution (for example during the entire refocusing phase); on the contrary, in the case where the pair cc, finds, in the subset MM (1) to MM (n), at least one classifier other than that defined by the "last acquisition" pointer, therefore in the where this pair is already known, these values a, p are stored in the corresponding STN modules. For example, if the pair o, p has been successfully classified by the subsets 1 and 2, these two values 1 and 2 are put on the LABEL bus L, and the classifiers of the associated unilinear modules STN (1), namely 811 and 812, are validated; these validate by retro-annotation the associated bi-linear STN modules, namely 801 and 802; finally the pair a, p is stored with the labels 1 and 2, in the subset MM (1)

  and in the subset MM (2), respectively.

  At the end of each treatment of the preceding type, the values RMAX and POSRMAX of each of the unilinear modules STN (1) affected are updated and a general test on all the unilinear modules 20 (STN1) already used makes it possible to recover the value

  The larger RMAX and its associated POSRMAX register is read and therefore corresponds to the most likely LABEL value, as discussed above with respect to the API calculation software.

  The contents of the associated bilinear histogram are then read and serve as anticipation of decision support as a dynamic recruiting agent in action, the treatment described previously in four phases being reduced to three (as indicated previously with reference to Figure 42), which accelerates the convergence for

the recognition of the object.

  Once the maximum resolution obtained, that is to say in particular at the time of complete refocusing, the calculation phase stops and the analysis phase of the results begins. This is characterized by reading the histogram memories of the bilinear modules STN (2) and unilinear STN (1), comparing the values read with a reference threshold, namely RMAX / 2, and the

  transcription in the associated 1-bit memory. The result of the 1-bit memory of the classifier of the bi-linear module is extended in order to accept adjacent pairs of ax, p, during subsequent tests. The final value of LABEL corresponds to the name of the object included in zone Z0. This value LABEL corresponds to so-called episodic storage.

  When processing on the L bus several LABEL values can be activated at the same time. The fact of cumulating these concidences by calculation of histogram with unilinear STN module 670 with multiclass extraction reveals new classes, CAT-1, CAT-2, CAT-3,. and these classes associating different values of LABEL between them, are the reflection of a categorization memory. Indeed, several objects can belong to the same category, for example: category of 15 faces including the labels of this or that person. This process

  and associated device brings up the two forms of memorization, namely, episodic and semantic. There is first knowledge of the event which in the course of successive learning enters categorizations, the knowledge of the 20 objects.

  According to one of the features of the invention, it is possible to solve the problem of the apparent variation in the size of the object under examination, for example when there is a relative approximation or distance between the object and the device of the object. observation, such as a camcorder. Achieving the resolution of invariance in

  size is illustrated in Figs. 54, 55a, 55b, 55c and 55d.

  FIG. 54 shows an assembly constituted by the same electronic elements as the assembly illustrated in FIG. 42, but the processing is performed on different input parameters 30. The basic assembly has been described with reference to FIG. 34, but, at the two input parameters x and y corresponding to the Cartesian coordinates, coming from the sequencer 9, the input parameters p and logarithm LD of the derivative distance are substituted. by the unit 400 of the initial parameter Dis distance, 35 from the sequencer 450, itself from the external unit 25 (not explained). The assembly of FIG. 54 uses a unit 450 corresponding to a sequencer, which outputs the various potential values LD and p during the time of the sequence and comprises three units, namely 400, 451 and 440. The unit 400 is constituted by a log calculator receiving as input the distance Dis and calculating LD, the logarithm of Dis which constitutes the output LD; the unit 451 is a classification unit outputting a binary signal Barpi, equivalent to the BarZi signal used previously, which is validated during the equality between the signals LD and P coming from the sequencer 450 and the values LDI and pi presented; finally the unit 440, which corresponds to the dynamic recruitment module M2 defined in FIG. 42, with instead: * The sectorization of the LD / p plane by the use of the variables LD and P coming from the sequencer 450, equivalent to the variables x and y in FIG. 42, transformed by equivalent units 420

  to 427 which each provide two axis transformations along Oi and Oj, such as units 150 and 151 previously described.

  This unit 420 outputs two signals 4200s and 4201S corresponding to the definition of a sector ZrO by the transformations of angle axes 00 and 01, the following units 421 to 427 ensure the sectorization of the plane defined by the variables LD and P . * The calculation of the invariance of the input data LDi / pi by the use of bilinear STN (2) modules 430 to 437, which receive, as input data, the signals 420is and 420js and which are controlled by the zone Z validation signals and the Barpi signal corresponding to the values

of entries LDi, pi.

  Figure 55d illustrates the result after treatment of dynamic recruitment STN modules during three consecutive sequences illustrated in Figure 55a, 55b and 55c. The units 400 and 451 continue, only the unit 440 is replaced by the units 401 and 402. The unit 401, which corresponds to the unit 150 of FIG. 34, receives the sequence LD and p coming from the unit 450. and on its control input ei corresponding to pi; finally the unit 402 is a one-dimensional module STN (1) which receives as input p (signal 401s)

  according to 0, noted po is controlled by Barpi.

  The input value Dis, which is equal to the distance from the object to the sensor 2 (such as a camcorder), or 2 'such as a radar (FIGS. 1, 5 2, 4a, 4b, 4c, 5 17, 18), measured externally, for example by means of a laser associated with said sensor, is converted, as previously indicated, to its logarithmic value, the two parameters Dis and p being the essential input parameters of the mounting of the sensor. Figure 54. Instead of using a laser, the distance Dis lo can be measured for example by binocular convergence, using parallax, using the knowledge of the size

real object perceived.

  At the output of the dynamic recruitment module 440, there are, as illustrated in FIG. 42, groups of STN modules successively in play; as in the case illustrated in Figure 43a, corresponding to what is obtained with the assembly of Figure 42, is obtained with the assembly of Figure 54 sectorization illustrated in Figure 55a, on which we find the areas or sectors such as Zro, Zr1, Zr2, Zr3 ... Zr7. On the other hand, instead of using the BarZo center of gravity, it is around the point referenced po, LD 0 defined by Barpo that sectors Zro to Zr7 are arranged. Figures 55b and 55c correspond to Figures 43b and 43c respectively, but this time also, the centers of gravity such as BarZo and BarZ1 are replaced by the points referenced 25 po, LDo and pl, LD1 respectively Barpo and Barp1, still in zone Zr3. In fact, if FIGS. 43b and 43c are compared with FIGS. 55b and 55c, it can be seen that the x and y coordinate axes have been replaced by the coordinate axes p and LD. At the end of the sequence, FIG. 55c, the calculated values ALD, 0, Cp 'are written in registers 1070, 1071 and 1072 respectively, and only

  there remains a unilinear STN module (1) shown in FIG. 55d.

  The sectorization, in the case of FIGS. 54 to 55c, is thus made from a first measurement po, LDo and from a second measurement made at a distance different from the first and providing p1, LD1 represented, as poLDo in FIG. 55b, an orientation being defined by a line pe perpendicular to the direction of the couple of points p0, LDo and p1, LD1, as illustrated in FIG. 55c, in which, in a direction perpendicular to p0, Cp is found which is the output of the one-dimensional module 5 STN (1) 402, the value of Cp 'corresponding to the projection of the straight line passing through the pair of points p0, LDo and pl, LD1, on the straight line of slope pe perpendicular to this direction of the couple of points. What is essential is that this value Cp 'is independent of the distance Dis, which corresponds to a result which does not depend on the size of the object observed, therefore on its relative position by

relative to the sensor that observes it.

  In FIG. 56, the result of the measurements made by the assembly of FIG. 54, the operation of which has been explained with reference to FIGS. 55a, 55b and 55c, is illustrated. In this FIG. 56, two planes are illustrated, namely a first plane having for coordinates p and LD and a second plane resulting from the first plane.

  by a rotation of angle 0 and having for coordinates p 'and D'.

  By way of example, measurements made for successive distances of 50 cm, 75 cm, 100 cm and 150 cm have been illustrated, the resulting points being substantially aligned and having as an ordinate value C p ', relative to the axis p. It is thus found that the module p 'has become practically invariant whatever the distance (within reasonable limits) of the object observed, so its

apparent size.

  Figure 57 illustrates an improvement of the assembly of Figure d. In this figure 57, we first find the unit 400 of Figure 55d, sequencing by LD and p signals with its input Dis and its output LDI, namely the logarithm of the distance Dis. This value LD constitutes with pi the parameters of the unit 410 which comprises all the units 401 and 402 of FIG. 55d and, in addition, the improvements constituted by the units 403, 404, 405, 406, 407 and 408. The output of the assembly 410 is constituted by the values Cp 'and LD' having the index i, whereas the mounting of the units 401 and 402 of FIG. 55d only included the output 35 Cp '. In fact, the assembly of FIG. 57 also incorporates a method of controlling the distance, this method being complete it can therefore use either the external measurement of the distance Dis if it is available, or the internal calculation thereof in the opposite case. For this purpose, a multiplexer 403 receives, on the one hand LDi, from the unit 400 and, on the other hand, the output of the adder 406 which delivers, as explained hereinafter, the calculated signal LD 'and choose one of these two input signals. If the value of the logarithm of the distance, namely LDi, is valid, the multiplexer 403 transmits this value, otherwise it transmits the calculated value LD ', this output signal being applied to the unit 401 similar to the unit 401. of Figure 55d; the choice between these two values is controlled by a control signal 403c, activated at least during the learning phase, from the unit 408 which receives the input signal

LDi, and who analyzes its validity.

  Calculation of the value LD 'is carried out when the external value LDi is valid, the unit 407 comparing the value of the incoming LD1 with that of LD' calculated previously; this unit 407 controls the gain of two amplifiers K1 404 and K2 405 which is equal to - tgO for the first and 1 / cosO for the second. An adder 406 adds the two values pi x tgO1 and Cp '/ cosO0, the

  sum of these two values being equal to LD '(in FIG. 57, the index i is added for the parameters p, 0 and LD', on the one hand, and Cp ', on the other hand). The unit 407 corrects the amplifier coefficients K1 and K2 to reduce the difference between the calculated value LD'1 and the input value LD1.

  The main advantage of the device of FIG. 57, with respect to the device of FIG. 55d, is that it can possibly dispense with a measurement of the distance Dis and hence a determination of LD1 from this distance, once that the object was spotted and learned and also able to size a new object among the known objects without having to measure the distance. So there is

  as many calculations Cp 'as of memorized modules.

  FIG. 58 corresponds to FIG. 47d, but with the integration of modules 410 of FIG. 57 which ensure the calculation of size invariance. All the M flow-through modules p or a of FIG. 47d are followed by associated modules 410 carrying Cp 'and LD'. The values LD 'make it possible to appreciate a volume aspect by their relative variations, as well as a distance by a calculation of

their average values.

  As far as the invariance is concerned, it has been explained so far how, by performing the calculations from BarZ0, the translation invariance is solved, whereas the unit 410 of FIG. of distance, to solve the problem of invariance in size. With regard to the invariance in rotation, the last invariance to be respected if one wants to determine an object independently of its distance and its position, both in translation and in rotation, the latter problem is solved by the

mounting of Figure 59.

  The assembly of this FIG. 59 firstly comprises a bilinear module STN (2) 800 receiving, on the one hand, Cp 'and, on the other hand, a' and an assembly 900, itself constituted by another module. unilinear STN (1) 901, a counter 902, a multiplexer

903 and a summator 904.

  The role of the assembly of FIG. 59 is to apply a double correction to the pair p, a: p is converted to Cp 'as indicated

  previously, while oa is transformed into a 'by unit 900.

  During the search phase, for each value of (a, the counter 902 starts at ac and increments at 2z radians, while the multiplexer 903 controls, by a signal initiating a search phase the transfer of the value from counter 902 on

  a summator 904 modulo 2 ir radians, which computes a value a '.

  When the calculated pair Cp ', ca' corresponds to the value already

  As known, the classification in the module STN 800 commands the validation of the module STN 901 and the current value of the counter is stored in the histogram of the latter module.

  At the end of the refocusing, or more generally when obtaining the maximum resolution, analysis of the histogram of the STN 901 module gives, as a result, the content of its POSRMAX corresponding to the average value cc '. The multiplexer 903 is switched and this new calculated value is transmitted to the summator 904. This makes it possible to recognize the object, the value Acc which

  out of the set 900 indicating the rotation of the object.

  Referring now to all of Figs. 60a and 60b, there is shown in Fig. 60a a portion of the anterior Fig. 58 feeding the triplets Cp ', a, LD' into the MM block of Fig. 53 enhanced by incorporation of the integration of the calculation of invariance in size and in rotation; in fact (FIG. 60b), at submodule M (0) of FIG. 53, a unilinear module STN (1) 920 which calculates the rotation Au and a unilinear module 10 STN (1) 921, which calculates the average distance, is integrated. LD ', the module 800 corresponding to the STN module 800 of Figures 53 and 59, while the module 810 of Figure 60b corresponds to the STN module 810 of

Figure 53.

  Fig. 61 schematically illustrates object recognition by increasing resolution and sensor rotation 2-5 or 2-5 ', by generating an egocentric landmark relative to and by this observer, by memorization of the organization of the shapes of the objects OB1 and OB2, the angles D1 and

D2 being equal.

  The rotation of the sensor actually makes first a decrease in the spatial resolution (period T1 of Figure 3); at the occurrence of a relatively high value of the signal Dif during rotation, a stop command of the rotation is issued by means (not shown) and the Gaussian increase process of the spatial resolution described above. above begins (period T3) of Figure 3); Moreover, the angle of rotation of the

  sensor is determined and stored.

  This rotation can be performed along two axes and this requires in principle the memorization of two angles, as well as the order of rotations along the two axes. However,

  choosing the axes of rotation of the sensor in the plane of Listing, the order of rotations along the two selected axes is indifferent and the stored information is then reduced just at the two angles of rotation, which is advantageous, because we can then 35 reconstitute a real egocentric landmark.

  Finally, in FIG. 62, there is illustrated what the sensor 2 (FIGS. 1 and 2) perceives and the result obtained by the method according to the invention: each object is * named by its label LABEL (namely 1, 2. ..), * positioned in distance by a value of LD (1, 2, 3 ...), * positioned in position by BarZ (1, 2, 3 ...), * positioned in rotation by Au. (1, 2, 3), and

  * characterized by its color C (1, 2, 3).

  Of course, the sensor perceives a part of the scene in the environment that depends on the position and the focal length thereof, which means that the notion of "object" also depends on these two parameters; as well as the sensor can transform into successive frames or sequences either a whole person, or his face, or a part of his face, the "object" being each time different.

  The method and the device according to the invention, of which the preferred embodiments have just been described, and which are generic tools applicable to other types of data and / or input parameters, make it possible to obtain, in the end Finally, the results indicated below result in a number of applications. A) The main results are as follows 1 With the method and the device described with reference to FIGS. 1 to 16, ie essentially with the whole of FIG. 1 or 2, implementing a digital signal (In particular a digital video signal with temporal sequences representative of successive views of an object, and spatial sub-sequences, representative of locations on the object, and realizing successively, on the one hand, after a reduction of the spatial resolution, a substantially Gaussian increase of this one, the reduction, as the increase, of the spatial resolution being carried out either by defocusing (FIG. 1) or by electronic filtering (FIG. 2) and, on the other hand, a spatio-temporal differentiation , with smoothing, of said increase, substantially Gaussian, between two successive sequences and for identical locations, which makes it possible to obtain a derived digital signal M 'showing successively hierarchical details more and more characteristic of the object (figures 13a-13b and 16). Thus, thanks to the generally abrupt decrease in the resolution and the increase, generally in steps, of the resolution (FIG. 3), the initial digital signal is replaced in which all the details were present at the same time. , without any 0o hierarchy, and therefore difficult to use to characterize the object they represent, by a derivative signal in which the details appear in descending order of importance, from the most characteristic of the object to the least characteristic of that that is, according to a hierarchy that can be materialized, by means of appropriate treatment, by a tree with a certain

number of filiations father -> son.

  With the more elaborate method and apparatus described with reference to FIGS. 17 and following, that is to say essentially - either with the whole of FIG. 17 implementing optical variation (by defocusing and refocusing), of the spatial resolution, or with the set of FIG. 18 with a variation by electronic filtering of this resolution, the aforementioned derivative signal M 'undergoes a processing consisting in forming, by means of one or more STN modules (of which the preferred structures are explained in relation to FIGS. 20, 23a-23b, 23c and 25), one or more representative histograms of the weight, that is to say quantities, of the different levels of the data contained in said derived signal M and thus in said initial digital signal (in particular the initial digital video signal), which allows first of all to determine the successive zones corresponding to the details appearing in order of importance. decreasingly in said desired signal and calculating the positions of these zones, with their centroid determined from these weights in their order of appearance, therefore of decreasing importance, and then advantageously to achieve a hierarchical representation in the form of 'a tree (figures 44c-44e to 49) with

father -> son relationships.

  Due to the determination of the centroids, characterizations of the observed object are carried out independently of any translation. Preferably, an angular sectorization (FIGS. 34 to 43c), by passing the Cartesian coordinates x and y at the polar coordinates p and cc (or p), makes it possible to determine relatively wide angular zones centered on the centroid of the first zone that has appeared. (the one with the most notable feature), then narrower angular sub-areas covering the relatively wide area, within which a new centroid has been found, these angular sub-areas also being centered on the center of gravity. said first zone, and

  to determine a first son relation to the father, defined by p in an absolute manner, independently of any rotation.

  This process is repeated for successive zones

following with their center of gravity.

  4 By applying the same type of sectorization, either with the parameters x, y then a, p, but with LD, p and D ', p', we can determine the invariance of p belonging to the object in relation to

  at the distance from the object to the sensor, regardless of the scale.

    In a preferred embodiment, the

  description of an object by its characteristics appearing in order of decreasing importance, by means of four successive sequences making it possible to calculate, by change of

  coordinates of x and y in p and a, elements relating to this object in the forward direction, that is to say by defining said object by its characteristics first the most important, then the least important; at the end of this four-step process, not only was the tree with the above-mentioned father-son filiations determined, but also a label or label defining

the examined object.

  It is then possible, in the presence of a new examined object, to quickly check whether this object substantially corresponds to the previously memorized object and, at the beginning of a possible recognition, the dynamic process can be realized more rapidly by suppressing the third sequence, so as to go directly from the second to the fourth sequence, with the examination of the identities between the elements of the object already stored io and the corresponding elements of the new object in this fourth sequence, become third sequence. It is found that there is an association of a label with the characteristics found, in descending order of importance, of the object, that is to say that to each of these characteristics of the object corresponds a portion 15 the label; Thus in the learning period, that is to say, the observation of a first object, there are cumulative characteristics of it in a label. When we examine a new object, that is to say, in recognition of the latter, to verify whether it corresponds to an object already analyzed and labeled, the emergence of an already known label makes it possible to extract the characteristics associated with this object.

  label to accelerate recognition.

  6 The rotation of the sensor according to two angles (which defocus in a period To of FIG. 3) makes it possible to determine the appearance of a first important characteristic (thanks to the high value of the signal Dif) of the observed object. . As the sensor stops in its rotational movement, the details appear, because of a refocusing, in descending order of importance and simultaneously this information concerning the

  details are compared to the information already in memory in the form of a label concerning previously examined objects, which makes it possible to accelerate the examination of the new object by possibly skipping steps in the recognition of the new object 35 and its comparison with an already known object .

  7 It is possible, ultimately, to obtain complete information on the object regardless of its rotation, translation and scale, for the reasons explained above. By applying, no longer to an object, but to an observed scene comprising several objects, the method according to the invention, it is possible to start from a representative zone ZO of the scene and its center of gravity BarZO to determine the location of the objects. different objects of the scene, with their ca, p associated with 1o BarZO, and to memorize a scene with its objects, in the form

  a label that will then determine if a new scene is observed or not by the initial scene labeled. Such a procedure can be used in particular in navigation, the principles of which will be explained below in the context of the applications of the invention.

  B - The method and the device according to the invention have many applications, some of which will be mentioned by way of non-limiting examples 1 - As described and illustrated, the recognition of a face, and therefore of a person, is one of the preferred applications of the invention. The perception of a face and the recovery of its associated labels make it possible to compare the labels of a new face with the labels of the faces already examined. The memorization of the face labels as and when they are examined can be performed in a calculator, a smart card, a magnetic identity card or a bar code, etc. From a stock of Such face label information makes it possible to check whether a new face, which has just been examined by the method and the device according to the invention, corresponds to labels already stored. It is thus possible to recognize an individual in particular for the purpose of identifying an individual wanted for various reasons or to allow an individual to access a particular

  determined area whose access is selective.

  Another application in the recognition of 35 faces is as follows: conventional type credit cards are perfected by adding, in their chip, the label or labels of the owner; when it comes before a cash machine and introduces its smart credit card with the label or labels of its face, the distributor 5 having a camcorder or a "webcam" is capable, after implementing the method and device according to the invention, to check whether the image of the face detected by the camcorder or the webcam corresponds to the label or labels inscribed in the chip of the credit card just introduced by the applicant of cash; this prohibits, in the event of a lost or stolen card, the use of it by someone other than its owner, since the face of the unauthorized user does not correspond to the label or labels which

  are stored in the chip of the credit card.

  The label or labels of a person may also be introduced into an identity card, a driving license, a

gray card, etc ...

  A similar application can be applied to mobile phones with a visual sensor; when first used by an authentic purchaser, the telephone will be able to determine the labels of the latter and possibly of one or more authorized persons; it will then be able to verify, by comparison of the labels stored with the user's labels, whether the current user is an unauthorized person in possession of the telephone as a result of a theft or loss; in this case, the telephone may transmit the label of the unauthorized person to the operator of the network on which the telephone depends.

  for all intents and purposes and possibly block the phone.

  We see finally that the method and the device according to

  The invention allows the identification of a person for a large number of applications.

  The invention can also be applied to the indexing of a photo library or image bank having a large number of photos. In a first phase, one or more labels are associated with each photo or image, for example a label corresponding to "cow" or "sunset". In a second phase, each label is associated with an alphanumeric reference which is stored in a "dictionary" next to the corresponding label and the label and reference number set.

  of the photo or image is stored in a memory.

  When a user wants to find pictures or pictures,

  among those stored, featuring a cow or a sunset, it introduces a keyword indicating cow or sunset.

  The dictionary of the library translates this keyword in alphanumeric reference, which makes it possible to find, in the memory, the _o number of the one or more photos containing the keyword. Of course when a user wants to find a picture or an image including several characteristics, for example a sunset and a

  cow, it will suffice to introduce the two corresponding keywords.

  Another application of the invention relates to navigation.

  For example, in the case of maritime navigation, a user of the method and of the device according to the invention, in order to be able to find their bearings, first examines with the sensor the environment of the place where it is located, in particular a port or a jagged coast, and this environment is labeled. Then, the sensor is moved along a nautical chart and, in case of identity, between the already memorized labels of the environment and the labels of an area of the nautical chart at the observed location, it can determine where is

his boat.

  Another application to navigation consists in locating a path, for example to find it when one wants to return to the starting point. Thus, as and when displacement, a device according to the invention allows to label successive positions with the corresponding environment; on return, the "navigator" tries to find, in the environment he observes with the sensor of the device, whether he recognizes a

place already visited one way.

  The invention will be defined by the following claims, in which, as in the above description, the term "object" includes not only the objects themselves,

  but also people or portions of people, especially faces, and animals, as well as possibly images or scenes (especially for navigation applications).

Claims (47)

  A method of operating a multiclass functional unit for calculating and processing a histogram for analyzing a parameter carried by a digital DATA (A) input signal in the form of a sequence Aij ... t, A ij ... t, A "ij ... t, ... binary numbers associated with synchronization signals making it possible to define a moment t of space and a position i, j ... in this space, The histogram may include a plurality of vertices and the process of producing a set of histogram characterization results including at least the maximum amplitude data pair of the histogram and the position of said maximum amplitude corresponding to the maximum peak of the histogram. histogram, characterized in that the histogram of the parameter is calculated as a function of a validation signal (VALIDATION) during a computation cycle, and a set of couples is determined during said calcu- lation cycle RMAXi amplitude and POSRMAXI position, the RMAX couples I and POSRMAXI being automatically ranked in descending order of vertex and stored in a functional memory with automatic material sorting of the functional unit   multiclass during said calculation cycle.   2. Method according to claim 1, characterized in that a set of results in the form of 25 k classification signals (CI1 ... Clk) with k greater than or equal to 1 is also output, each classification signal corresponding to a class in   report with a top of the histogram.   3. Method according to claim 1 or 2, characterized in that a multi-class functional unit is implemented which comprises an application programming interface (API), a memory M0 storing the ordered amplitudes RMAXj of the vertices, a memory M1 of vertex-ordered addresses POSRMAXI of vertices, memories M0 and M1 being grouped into a single functional table of amplitude and memory position pairs (RMAXo / POSRMAXo; RMAX1 / POSRMAX1; RMAX2 / POSRMAX2, ... RMAXi / POSRMAXi. ..) vertices of the histogram in decreasing order of magnitude, the functional table (600) performing an automatic material sorting of classes during the computation cycle, the multiclass functional unit 5 further comprising a memory M2 for storing the serial number of the class, a class threshold memory M3   and an RC register of number of classes.   4. Method according to claim 3 characterized in that the API makes it possible to perform: an initialization cycle resetting the memories M0, M1, M2, M3 and the register; a calculation cycle for determining and classifying automatically in the functional table the pairs of values, - a class update cycle, i5 the update cycle including (B) - a class search step with class labeling in the memory M2 and storing thresholds corresponding in the memory M3, the number of classes thus determined being stored in the register RC, (C) - a step of validating the classes of the memory M2 by comparison of the value of the memory M0 to the address of the class considered of M2 with the threshold of M3 corresponding to the class considered.   5. Method according to claim 3 or 4, characterized in that for each class, the number of NBPTSi points is determined and memorized in a memory M4.   the histogram corresponding to said class.   6. Method according to claim 5, characterized in that the average position POSMOYi of said parameter is determined and memorized further for each class in a memory M5. said class.   7. The method as claimed in claim 6, characterized in that a barycentric unit for producing a barycentre output signal with a first binary state 35 is also used when the parameter corresponds to the average POSMOYi position and with a second binary state otherwise for the class considered.   8. Process according to any one of the claims   characterized in that a selection means (1162) is used for selecting for a given class a threshold criterion as a function of at least one of the following values: the value of the amplitude of the vertex of the histogram, - a THRESHOLD threshold value supplied to the unit, - a number of NBPTS points of the histogram, each of the classes of the classification signal being generated in relation to a set of parameter values for which   the amplitude of the histogram is greater than said threshold criterion.   9. Process according to claim 8, characterized in that for a given class, the criterion is selected from RMAX / 2. THRESHOLD, NBPTS / THRESHOLD.   10. Process according to any one of the claims   characterized in that the parameter analyzed by the multiclass functional unit is complex and obtained by combining at least two elementary parameters, each of the binary numbers of the DATA input signal (A1, A2, A3 ... Ap) support of the complex parameter (AIA2A3 ... Ap) comprising P fields each corresponding to an elementary parameter A1, A2, A3, Ap.   11. Method according to any one of claims 2 to 10   characterized in that the multi-class functional unit is implemented in a histogram computation and processing module, said STN, the output signals forming a retro-annotation being sent on a feedback bus (114) and the signal of 30 validation (VALIDATION) involving the retroannotation.   12. Method according to claim 11, characterized in that in an object recognition system comprising a set of calculation modules and histogram processing, at least two STN modules with a multiclass functional unit are used, a first module (660) (664) operating in a time domain TD, determining at least one class and recruiting for said class at least one second module (661, 662,   663) (665, 666, 667) operating in a SD space domain.   13. The method as claimed in claim 12, characterized in that the module operating in the TD time domain receives a   MVT speed setting or L / T / S color.   14. Process according to any one of the claims   characterized in that it is implemented in an object recognition system comprising a set of STN computation modules and histogram processing by sectoring with at least one multi-class functional unit module, the modules determining zones. and in that a specific area is divided into several angular sectors centered on the corresponding centroid of the area and that one looks in which sector among the sectors, a new barycenter appears and that one division into several sub-sectors, which process may continue in order to   to continue to progressively sharpen sectorisation.   15. Method according to claim 14, characterized in that it is implemented in at least one STN module with at least two orientation units p13 of axes (150, 151) input for rotation of axes of reference of at least two Cartesian coordinates of input parameters, the module (s) further determining the barycenter for the input parameters, and in the system determining a first space Z1 having a barycenter BarZ1 by an association of a monolinear module (296) processing a first parameter and a second bilinear module (297) processing the coordinates, a second association (298) (299) determining within said first space Z1 the appearance of a second barycentre BarZ + 1, said first space being divided into distinct regularly distributed angular sectors (Zrio, Zr1il, Zr! 2 ...), each of the sectors being treated by a bilinear sector module (300, 301 ... 307) receiving Z1, BarZ1 and also the signal of the second barycenter BarZ1 + 1, the bilinear sector module corresponding to the second barycenter BarZi + 1 being connected to a set of bilinear modules sector of subsequent rank to split the sector having the second barycenter BarZ, + 1 in distinct angular sectors regularly distributed (Zra, Zrb, Zrc ...) in order to progressively refine the sectorization and further determine at least an angle α (1071) of reference axes substantially perpendicular to the direction of the straight line uniting the two barycentres BarZi and BarZi + l and a module pj (1070) distance between the two barycentres BarZ1 and BarZi + 1 along said straight line, lo the reference according to the angle a to obtain a translation invariance. 16. The method according to claim 15, characterized in that the object can be observed at different distances and the system is furthermore implemented in at least one size invariance unit (450) (451). (410), said size invariance unit receiving at least an input, on the one hand, a value of the logarithm (400) of a distance LD between a reference point and at least one point of the object and, on the other hand, the distance module pj between two centroids BarZ1 and BarZ1 + 1, said unit determining at least one projection value Cp '(1072) substantially constant and corresponding to an angle φ (1071) of rotation with respect to referential p and LD.   17. The method as claimed in claim 16, characterized in that distance control means are used in the size invariance unit (410) enabling the choice to be made using an external distance measurement. or the internal determination of the distance.   18. A method according to any one of claims 15 to   17 characterized in that the object can be viewed at different angles and that the system is further operated with at least one rotation correction unit (900), said rotation correction unit for correcting the reference axis ac angle value with respect to a pair of ac values,   p of angle and module previously determined.   19. Process according to any one of Claims 14 to   Characterized in that it is implemented with at least one parameter marker transformation unit by rotation of angle θ, the marker being at least two dimensions of parameters, the rotation in the case of a two-dimensional reference for pixel polar coordinate parameters X, Y, corresponding to the following matrix operation [X] [Scos 0 sin 0 x X Lyj Lsin 0 -cos 0 y   A process according to any one of claims 14 to   Characterized in that a plurality of centroids of areas appearing successively with respect to a first one of an initial zone is determined over time and the centroid coordinates are stored in relation to a label in the form of recognition.   21. Process according to claim 20, characterized in that   to determine whether or not a new observed object corresponds to a previously stored label object, the recognition data of the new object is determined and compared to previously stored label recognition data.   22. The method of claim 20 or 21 characterized in that   furthermore, the label recognition data is associated with an angle of rotation Ata and a mean distance LD 'by means (920) (921) for determining said rotation angle Δ 0 and said average distance LD'.   23. A method according to any one of claims 20 to   22 characterized in that the label recognition data are analyzed by a histogram calculation and processing module (670) capable of determining and storing a set of categorization data of said labels.   24. Multi-class computing functional unit and histogram processing for analysis of a parameter carried by a digital DATA (A) input signal in the form of a sequence Aij .. t, A'ij ... t, A "j ... t, ... binary numbers associated with synchronization signals making it possible to define a moment t of space and a position i, j ... in this space, the histogram possibly comprising several vertices and the processing of producing a set of histogram characterization results comprising at least the maximum amplitude data pair of the histogram and the position of said maximum amplitude corresponding to the maximum peak of the histogram, characterized in that it comprises means making it possible to calculate the histogram of the parameter as a function of a validation signal (VALIDATION) during a computation cycle, and to determine during said computation cycle a set of amplitude pairs RMAXI and position POSRMAXi, the RMAXi couples and POSRMAXI are automatically ranked in descending order of vertex and stored in an automatic material sort functional memory of the multi-class functional unit during said calculation cycle.   25. Unit according to claim 24, characterized in that the means furthermore make it possible to output a set of results in the form of k classification signals (CI1 ... Clk) with 20 k greater than or equal to 1, each signal of classification corresponding to a class related to a top of the histogram. 26. Unit according to claim 24 or 25, characterized in that it comprises an application programming interface (API), a memory M0 storing the ordered amplitudes RMAX of the vertices, a memory M1 of amplitude-oriented addresses POSRMAX, vertices, memories M0 and M1 being grouped into a single functional table (600) of amplitude and memory position pairs (RMAXo / POSRMAXo; RMAX1 / POSRMAX1; RMAX2 / POSRMAX2, ... RMAXI / POSRMAXi ...) of the vertices of the histogram in decreasing order of magnitude, the functional table (600) performing an automatic material sorting of the classes during the calculation cycle, the multi-class functional unit further comprising a memory M2 for storing the serial number of the class, a class threshold memory M3   and an RC register of number of classes.   27. Unit according to claim 26 characterized in that the API makes it possible to carry out an initialization cycle resetting the memories MO, M1, M2, M3 and the register, a calculation cycle for automatically determining and classifying. in the functional table the pairs of values, - a class update cycle, the update cycle comprising (B) - a class search step with class labeling in the memory M2 and storage of the corresponding thresholds in the memory M3, the number of classes thus determined being stored in the register RC, (C) - a step of validating the classes of the memory M2 by comparison of the value of the memory M0 with the address of the considered class of M2 with the threshold of M3 corresponding to the class considered.   28. The unit as claimed in claim 26 or 27, characterized in that it makes it possible to determine and memorize furthermore for each class, in a memory M4, the number of points NBPTSi of   the histogram corresponding to said class.   29. The unit as claimed in claim 28, characterized in that it makes it possible to determine and memorize further, for each class, in a memory M5, the average POSMOYI position of said parameter in said class.   30. Unit according to claim 29, characterized in that it further comprises a barycentric unit intended to produce a barycenter output signal with a first binary state when the parameter corresponds to the average position POSMOYi and with a second binary state in the opposite case for the class considered.   31. Unit according to any one of the claims   previous characterized in that it comprises a means (1162) of selection for selecting for a given class a threshold criterion according to at least one of the following values: the value of the amplitude of the top of the histogram a threshold value supplied to the unit, a number of NBPTS points of the histogram, each class of the classification signal being generated in relation to a set of parameter values for which   the amplitude of the histogram is greater than said threshold criterion.   32. Unit according to claim 31, characterized in that 10 for a given class, the criterion is selected from RMAX / 2, THRESHOLD, NBPTS / THRESHOLD.   33. Unit according to any one of the claims   previous characterized in that the parameter analyzed is complex and obtained by combining at least two elementary parameters, each of the binary numbers of the input signal DATA (A1, A2, A3 ... Ap) support of the complex parameter (A1A2A3 ... Ap) comprising P fields each corresponding to one   elementary parameter A1, A2, A3, ..., Ap.   34. Unit according to any one of claims 25 to 33, characterized in that it is in a calculation module and   histogram processing, said STN, the output signals forming a retroannotation being sent on a feedback bus (114) and the validation signal (VALIDATION) involving the retroannotation. 35. Unit according to claim 34, characterized in that it is in an object recognition system comprising a set of calculation and histogram processing modules comprising at least two multi-class functional unit STN modules, a first module (660). ) (664) operating in a time domain TD, determining at least one class and recruiting for said class at least one second module (661, 662,   663) (665, 666, 667) operating in a SD space domain.   36. Unit according to claim 35 characterized in that the   TD time domain module receives a MVT or L / T / S color parameter.   37. Unit according to any one of the claims   characterized in that it is in an object recognition system comprising a set of STN modules for calculating and histogram processing by sectoring with at least one multi-class functional unit module, and that the modules make it possible to determine zones and centers of gravity and to divide a given zone into several angular sectors centered on the corresponding center of the zone and to search in which sector among the sectors, a new center of gravity appears and to divide the said sector into several sub-sectors, the said process being continue in order to continue to   gradually refine the sectorization.   38. Unit according to claim 37 characterized in that it is in the system comprising at least one STN module with at least two orientation units poa p3 axes (150, 151) input for rotation axes of reference of at least two Cartesian coordinates of input parameters, the module (s) further determining the barycenter for input parameters, and in the system determining a first Z space, having a BarZi center of gravity by an association of a monolinear module (296) processing a first parameter and a second bilinear module (297) processing the coordinates, a second association (298) (299) determining within said first space Zi the appearance of a second BarZ1 + 1 barycenter, said first space 25 being divided into distinct regularly distributed angular sectors (Zrio, Zr1il, Zri2 ...), each of the sectors being treated by a bilinear sector module (300, 301 ... 307) receiving Zi , BarZ1 and equal the signal of the second barycenter BarZ + 1 ,. the bilinear sectoral module corresponding to the second barycenter BarZl + 1 being connected to a set of bilinear modules of subsequent rank sector making it possible to split the sector having the second barycenter BarZi + 1 into distinct regularly distributed angular sectors (Zra, Zrb, Zrc ...) in order to progressively sharpen the sectorization and furthermore determine at least an angle α (1071) of reference axes substantially perpendicular to the direction of the straight line uniting the two centroids BarZ1 and BarZl + and a module pj ( 1070) of distance between the two centroids BarZi and BarZ [+ 1 along said straight line, the reference along the angle ax making it possible to obtain translation invariance. 39. Unit according to claim 38 characterized in that the object can be observed at different distances and the system further comprises at least one size invariance unit (450) (451) (410), said unit size invariance receiving at least 10 input, on the one hand, a value of the logarithm (400) of a distance LD between a reference point and at least one point of the object and, on the other hand, the a module pj of distance between two centroids BarZi and BarZ1 + 1, said unit determining at least a substantially constant projection value Cp '(1072) and corresponding to an angle θ (1071) of rotation with respect to referential p and LD.   40. The unit as claimed in claim 39, characterized in that the size invariance unit (410) comprises distance control means allowing the choice to use an external distance measurement or the internal determination of the distance. distance.   41. Unit according to any one of claims 38 to 40   characterized in that the object can be viewed at different angles and the system further comprises at least one rotation correction unit (900), said rotation correction unit for correcting the angle value ac of reference axis with respect to a pair of values a, p of angle and module previously determined.   42. Unit according to any one of claims 38 to 41 characterized in that it is associated with at least one unit of   transformation of parameter mark by angle rotation 0, the mark being at least on two parameter dimensions, rotation in the case of a two-dimensional mark for pixel polar coordinate parameters X, Y, corresponding to following matrix operation: FXl cos 0 sin 0 x ---- X lY- Lsin 0 -cos 0 Y   43 Unit according to any one of claims 38 to 42, characterized in that in the system means   to determine, over time, several centroids of zones appearing successively in relation to a first one of an initial zone and to memorize the centroid coordinates in relation to   with a label in the form of recognition data.   44. Unit according to claim 43 characterized in that   to determine whether or not a new observed object corresponds to a label object previously stored, the system makes it possible to determine the recognition data of the new object and to compare it with previously stored label recognition data.   45. Unit according to claim 43 or 44, characterized in that   that in addition to the recognition data of the label are associated an angle of rotation Ac and a mean distance LD 'by means (920) (921) for determining said angle of rotation Aoa and said average distance LD'.   46. Unit according to any one of claims 37 to 45   characterized in that the system enables the label recognition data to be analyzed by a histogram calculation and processing module (670) capable of determining and storing a categorization data set of said labels.   47. Application of the process of any one of   Claims 1 to 23 with a multiclass functional unit according to one of Claims 24 to 46 in a system for the perception, recognition and localization of an object in its environment from a signal   digital input consisting of a succession of sequences of successive views of the object in its environment and thus falling within the time domain, each of said sequences being constituted by a succession of subsequences, each representative of locations arranged one to the sequence of the other in said sequences and therefore falling within the spatial domain, characterized in that a temporal variation of the spatial resolution of the object in said digital input signal is carried out during a period of several sequences, the variation comprising a substantially Gaussian increase phase of the resolution from a reduced value to an optimal base value, a differentiation is further carried out, smoothing between two successive sequences of said substantially Gaussian increase in the resolution, in order to obtain a derived digital signal representative of the variability of the Gaussian difference between these two sequences when the difference, in absolute value, for each same spatial location of said derived signal, exceeds a threshold, and deduced from said derived signal, by comparison in the successive sequences between the values of this derived signal corresponding to locations identical in the subséquences, hierarchical details of the object, forming at least two histograms of said derived digital signal, at least one of which is relative to the numerical magnitude of said signal in the various locations, which provides information relating to the details features of the object, and at least one other of which relates to location of locations in said signal, thereby providing location information said details.