WO2016151103A1

WO2016151103A1 - Method for counting objects in a predetermined spatial area

Info

Publication number: WO2016151103A1
Application number: PCT/EP2016/056630
Authority: WO
Inventors: Antoine FAGETTE; Stephen CHAI KHEH CHEW; Nicolas COURTY; Thomas CORPETTI
Original assignee: Thales; Universite De Bretagne Sud; Centre National De La Recherche Scientifique
Priority date: 2015-03-24
Filing date: 2016-03-24
Publication date: 2016-09-29
Also published as: FR3034229B1; SA517390016B1; SG11201707853WA; EP3274914A1; FR3034229A1

Abstract

The invention relates to a method for counting objects in a predetermined spatial area, the method comprising: a first unsupervised learning stage, and a second stage of counting in the area, which includes the steps of: acquiring, by the detector (30), an image comprising the predetermined area, and estimating the spatial density of objects in the predetermined area by applying the law to the acquired image.

Description

Method of counting objects in a predetermined spatial area

The present invention relates to a method of counting objects in a predetermined spatial area. The present invention also relates to a computer program product and an associated information carrier. The present invention also relates to a detection device adapted to implement the counting method.

With the increasing use of video surveillance in public places to enhance security, many video analysis techniques provided by cameras have been developed.

It is particularly known from EP 2 704 060 A2 a method for estimating the size of a crowd from an image, the method comprising the steps of determining a candidate to obtain the most informative data point using an active learning algorithm. The method also includes a step of receiving an annotation from a user identifying the data of the most informative data point. The method also includes a step of applying a regression function to the most informative point determined and the associated notation to minimize a loss on the annotated data. The method includes, finally, a step of constructing a regression model using the user annotated data and the deduced annotated data.

Document US 2007/0031005 A1 describes an optical system comprising a camera and a processor comprising a video analysis software. The software is adapted to estimate a baseline frame representative of a background, estimate geometric parameters to represent a scale of object variation in the given frame, obtain a change detection map to distinguish the background from object in the given frame and combine the change detection map with the geometric parameters to obtain a measure of congestion in the given frame.

However, as the spatial density of objects increases, the aforementioned methods provide unreliable results in determining the spatial density of objects in the provided image.

There is therefore a need for an object determination method in a predetermined spatial area to obtain more reliable estimates, especially when the spatial density of objects is large.

For this, there is provided a method of counting objects in a predetermined spatial area. The method includes a first non-learning phase supervised method comprising the steps of selecting a plurality of discrete spatial densities of objects in the predetermined spatial area, providing parameters relating to a detector adapted to acquire at least one image, the image comprising the predetermined spatial area, the parameters provided comprising at least one parameter relating to the focal length of the detector, a parameter relating to the altitude of the detector and a parameter relating to the inclination of the detector relative to the predetermined spatial zone, of generating a database comprising a a plurality of images for each selected spatial density, the generation step using the parameters provided at the providing step, and determining a law for estimating the spatial density in the predetermined spatial area using the base d generated learning. The method also comprises a second enumeration phase in the predetermined spatial zone comprising the steps of acquiring an image comprising the spatial zone predetermined by the detector, and estimating the spatial density of the object in the predetermined spatial zone by application of the law in the acquired image.

According to particular embodiments, the counting method comprises one or more of the following characteristics, taken in isolation or in any technically possible combination:

the generation step comprises a step of correcting the perspective defect related to the detector.

the method comprises a segmentation step of an image for determining the pixels containing an object.

the determination step is implemented by learning, the learning comprising an extraction of characteristics, the extracted characteristics being also used in the estimation step.

the determination step is implemented using a carrier vector regression.

in the supply step, a parameter relating to the detector is an intrinsic characteristic of the detector.

in the supply step, a parameter relating to the position of the detector and / or an orientation parameter of the detector is provided.

There is also described a computer program product comprising software instructions, the software instructions implementing a method as described above, when the software instructions are executed by a computer.

It is also proposed an information carrier on which is stored a computer program product as described above. It is also proposed a monitoring device for counting the objects in a predetermined spatial area comprising a detector adapted to acquire at least one image, the image comprising the predetermined spatial zone, the detector having parameters, a parameter being relative to the focal length of the detector, a parameter being relative to the altitude of the detector and a parameter being relative to the inclination of the detector relative to the predetermined spatial zone. The device comprises a processing module adapted to count the objects in the predetermined spatial area. The processing module is able to choose a plurality of distinct spatial density in the predetermined spatial zone, to receive parameters relating to the detector, the parameters received comprising at least one parameter relating to the focal length of the detector, a parameter relating to the altitude of the detector and a parameter relating to the inclination of the detector with respect to the predetermined spatial zone, generating a plurality of images for each selected spatial density using the parameters received, determining a law making it possible to estimate the spatial density in the spatial zone predetermined by using as a basis of learning the generated learning base and estimate the spatial density of object in the predetermined area by applying the law to the acquired image.

Other features and advantages of the invention will appear on reading the following description of embodiments of the invention, given by way of example only and with reference to the drawings which are:

FIG. 1, a schematic view of an example of a processing module enabling the implementation of a method for counting objects in a predetermined spatial zone,

FIG. 2, an example of an image generated to constitute a learning base, and

- Figure 3, a flow chart of an example of implementation of a method of counting objects.

A processing module 10 and a computer program product 12 are shown in FIG.

The interaction of the computer program product 12 with the processing module 10 makes it possible to implement a method of counting objects in a predetermined spatial zone.

The processing module 10 is preferably a computer.

More generally, the processing module 10 is an electronic calculator able to manipulate and / or transform data represented as electronic or physical quantities in registers of the processing module 10 and / or memories in other similar data corresponding to physical data in the memories, registers or other types of display, transmission or storage device.

The processing module 10 comprises a processor 14 comprising a data processing unit 16, memories 18 and an information carrier reader 20. The processing module 10 also comprises a keyboard 22 and a display unit 24.

The computer program product 12 includes an information carrier 20. The information carrier 20 is a readable medium. An information carrier 20 is a support readable by the processing module 10, usually by the data processing unit 14. The information carrier 20 is a medium adapted to memorize electronic instructions and capable of being coupled to a bus of a computer system.

By way of example, the information carrier 20 is a diskette or floppy disk ("floppy disk"), an optical disk, a CD-ROM a magneto-optical disk, a ROM memory, a memory RAM, an EPROM memory, an EEPROM memory, a magnetic card or an optical card.

On the information carrier 20 is stored a computer program including program instructions.

The computer program is loadable on the data processing unit 14 and is adapted to be used to carry out the implementation of the method of counting objects in the predetermined spatial area when the computer program is implemented on the treatment module 10.

The operation of the processing module 10 interacting with the computer program product 12 is now described with reference to FIGS. 2 and 3 which schematically illustrate an exemplary implementation of the method of counting objects.

By the term "object" is meant a countable element. In the context of the invention, the object is observable, that is to say that a detector 30 as shown in Figure 3 is able to observe the object in the predetermined area.

Such a detector 30 is, by way of illustration, a camera 30.

Such a camera 30 is characterized by parameters P.

The detector 30 has thus intrinsic parameters P _in trinsèque, that is to say not dependent on the environment, as the focal detector 30 or the size of the image acquired by the detector 30.

The intrinsic parameters P _in tnnsque are parameters specific to the camera 30. The detector 30 also has extrinsic extrinsic parameters P. As an example, a parameter relating to the position _P.sub.0 of the camera 30 depending on the environment in question is an extrinsic parameter Pextnnsec-

Preferably, a position parameter P _poSition of the camera 30 is the altitude of the camera 30.

In another example, a position parameter P _poSition of the camera 30 is the distance of the camera 30 from the predetermined spatial area.

Usually, the distance from the camera 30 to the predetermined spatial area is defined as the distance between the plane of the detectors of the camera 30 and the point of the predetermined spatial area which is closest to the camera 30.

According to yet another example, assuming that an orthonormal coordinate system (X, Y, Z) is defined for the space to be monitored, other position parameters P _position of the camera 30 are the position along the X axis and the position along the Y axis.

Similarly, a parameter related to the orientation of the camera 30 and depending on the mechanical system holding the camera 30 is an extrinsic parameter Pextnnsque-

Preferably, the orientation parameter Guidance of the camera 30 is the angle of inclination of the axis of view of the camera 30 relative to the horizontal. Such an angle is generally referred to as the pitch angle.

In addition, other Orientation orientation parameters are taken into account. Another orientation parameter of the camera 30 is, for example, the angular orientation of the direction of the optical axis of the camera 30 relative to an absolute reference. The direction of the optical axis of the camera 30 is usually called the heading angle.

The inclination of the camera 30 around its axis of view, also known as the roll angle, is another example of orientation parameter Guidance of the camera 30.

It should be noted that the orientation parameters of the camera 30 are also expressible in the form of Euler angles, namely the angles of precession, nutation and own rotation. The nutation angle is the generally preferred parameter.

In the following, it is considered that the objects are persons and that the detector 30 is a CCTV camera. In such a case, the predetermined spatial area is a space to be monitored.

In another example, the objects are objects of biological interest observed with the aid of a detector 30 and an optics performing a microscope function.

According to yet another example, the objects are carbon nanotubes and the detector is in this case an electron microscope. More generally, it is possible to apply the method in question to any type of object whose behavior is simulable in a realistic manner.

The predetermined spatial area ZS generally depends on the object.

By way of illustration, when the object under consideration is an object of biological interest, the predetermined spatial zone ZS is a part of the sample observed using the microscope.

In the context of the example described, when the object is a person, the predetermined spatial area ZS is a part of the environment, for example, a street portion. This is clearly shown in the image of Figure 2.

In such a case, the camera 30 is suitable for acquiring an image of the predetermined spatial area ZS, typically the camera 30 is suitable for acquiring an image of the street. In addition, the enumeration method is a method of enumerating the crowd in an environment.

As schematically represented in FIG. 3, the method comprises two successive phases P1 and P2. The first phase P1 corresponds to the upper part of FIG. 3 (delimited by the dotted lines 32) while the second phase P2 corresponds to the lower part of FIG. 3. The first phase P1 is a non-learning phase. supervised while the second phase P2 is a counting phase in the predetermined spatial zone ZS.

The first phase P1 unsupervised learning aims to obtain a law for estimating the spatial density of people in the predetermined area.

By the term "unsupervised learning", it is understood that the learning data is generated automatically.

In such a context, the law is a link between features of an image and a value of the spatial density of people in the predetermined spatial area ZS.

The law thus makes it possible to determine characteristics involved in increasing the spatial density of people.

During the first learning phase P1, parameters of the detector 30 are used to generate a learning base automatically and, from the training base, a regression model is learned by the processing module 10 .

The second phase P2 of enumeration in the predetermined spatial zone ZS seeks to apply the law to real cases.

By the term "real case" is meant images acquired by the camera 30 and not generated automatically as in the case of the first phase P1 unsupervised learning. More specifically, the video stream from the camera 30 is collected, and characteristics are extracted from the video stream to obtain the spatial density over the entire image thanks to the regression model learned during the implementation of the first one. P1 learning phase.

The implementation of the method thus makes it possible to obtain the enumeration of persons in the predetermined spatial zone ZS.

The first unsupervised learning phase P1 comprises four steps: a selection step S100, a supply step S102, a generation step S104 and a determination step S106.

At the choice step S100, a plurality of spatially distinct density of people in the predetermined spatial area ZS is selected.

According to one embodiment, each spatial density is equidistributed between a zero spatial density and a high spatial density.

In general, a spatial density is considered high depending on the context and in particular the conditions of the shooting.

Usually, a spatial density greater than or equal to 50 people in the predetermined area is considered a high spatial density.

At the end of the choice step S100, a set of spatial density has thus been obtained.

According to one embodiment, the choice step S100 is implemented for several distinct areas of the image.

At the supply step S102, parameters relating to the camera 30 are provided. The parameters provided comprise at least one parameter relating to the focal length of the detector 30, a parameter relating to the altitude of the detector 30, a parameter relating to the inclination of the detector 30 relative to the predetermined spatial area ZS and a parameter relating to the position of the main point.

A parameter relating to the position of the main point is, for example, the position of the optical center relative to the center of the image.

In one example, a parameter relating to the camera 30 is an intrinsic parameter P _in trinsèque-

For example, the focal length of the camera 30 or the number of pixels of the camera 30 is provided.

In another example, the parameter relating to the camera 30 is at least one of the group consisting of a position parameter P _position and an orientation parameter Orientation of the camera 30 relative to the environment. At the end of the supply step S102, parameters relating to the camera 30 are obtained. Among these parameters, there is at least one parameter relating to the focal length of the detector 30, a parameter relating to the altitude of the camera. detector 30, a parameter relating to the inclination of the detector 30 with respect to the predetermined spatial zone ZS and a parameter relating to the position of the main point.

In the generation step S104, a learning base is generated.

For this, in the generation step S104, both a plurality of images are generated and an annotation element corresponds to the terrain truth. The term "ground truth" is more often referred to as "ground-truth".

The plurality of images is generated for each spatial density selected at the choice step S100.

By "generation", it is to be understood that the processing module 10 generates synthetic images by using the parameter relating to the camera 30. The synthetic images are not images coming from the camera 30. These are images generated as if they were taken by the camera 30 but which are synthetic images.

Preferably, the pedestrian positions comprising the synthetic images are calculated to correspond to a spatial density field in a three-dimensional environment included in an area comprising the field of view of the camera 30 and its immediate vicinity to avoid edge effects.

For example, to parameterize the spatial density field, a spatial density gradient is created ranging from a maximum spatial density to a minimum spatial density across the image. The gradient is oriented along lines of constant depth in the image, so that for a given depth, the spatial density ranges are from very dense to very rare. This is consistent with the assumption that for a given depth, the appearance of the constant spatial density crowd is about the same along a line of constant depth.

Alternatively, a mixture of Gaussian functions randomly positioned in the predetermined spatial area ZS with a random standard deviation is used. Such a mixture makes it possible to generate an image close to the crowds obtained in reality.

An approach allowing to implement the computation of the ground truth is to define a function of density of space of reality ground, called function F. To calculate the function F for each pixel of a generated image, the function of density of space of reality Field F is defined by the following equation: or :

- P is the list of positions in the image where pedestrians are,

N (p, P, σ ² ) is a two-dimensional Gaussian nucleus evaluated at the pixel p with the average position P, and

- σ is a coefficient.

The position P of a pedestrian in the image is calculated by using a parameter of the camera 30 by projecting into the image a point situated at the position of the pedestrian considered and at the altitude corresponding to half the height of the camera. a human. Typically, it is considered that the average height of a human being is 1.70 meters (m).

Similarly, the coefficient σ is calculated as one-third of one-half of the apparent height of a human being at position P in the image. Therefore, the value taken by the coefficient σ for the pedestrian P depends on its location in the image. Usually, the coefficient a is of the order of a few pixels.

The calculation of the ground truth is implemented automatically and avoids the annotation by a human being. Such an operation is therefore more precise since the operation is performed automatically.

At the end of the generation step S104, there is thus obtained a database that can serve as a training database. The database contains a set of synthetic images to which, for each image, is associated a ground truth.

In the determination step S106, a law is determined.

The law makes it possible to estimate the spatial density in the predetermined spatial zone ZS.

For this purpose, it is used as a learning base for the learning base generated in the generation step S104. As explained above, it should be noted that the learning base is generated without implementing a manual annotation by an operator.

According to the example of FIG. 3, the determination step S106 is implemented by four successive steps: a correction step S106a of the perspective, a segmentation step S106b of the crowd, an extraction step S106c of the characteristics and a learning step S106d.

In the correction step S106a, it is taken into account that in an image, the objects close to the detector 30 appear larger than the objects that are farther away. Such a perspective effect is taken into account when analyzing the characteristics to be extracted from the image because a characteristic extracted from an object in the foreground represents a portion of the object smaller than a characteristic extracted which is extracted further in the image. In the correction step S1 06a, two successive steps are implemented, namely a step of normalizing the characteristics by using a perspective normalization map and a step of dividing the image into quasi-deep depth bands. constant.

The normalization of the features is implemented using a perspective normalization map in which a pixel weight is related to the expected depth for the object that generated the pixel. A weighting of weights is applied for distant objects. Weighting increases the weight considered if the object is farther than an object in the foreground.

The perspective normalization map is generated using the camera settings 30.

More precisely, the correction factor is obtained by orthorectification. Thus, instead of using two-dimensional parameters such as height or width, it is better to take into account the three-dimensional nature of a human being. For this, the variation in appearance of the three-dimensional volume occupied by the pedestrian. From this is deduced a correction factor of the perspective W (p) for the pixel p.

By way of example, the correction factor of the perspective W (p) is obtained by means of the following formula:

max (A ₁ () + A ₂ (p))

W (p) = where:

• the human being is modeled by a cylinder with a circular base,

• Ai (p) is the apparent area of the lateral area, and

• A ₂ (p) is the apparent area of the upper circular base.

The map thus obtained is used to normalize the characteristics by applying the weights calculated to each of the pixels that are in the predetermined spatial area ZS if the characteristic is a surface-like characteristic and by applying the square root of the weights to the characteristics related to edge effects if the characteristic is a contour type characteristic.

For the segmentation in bands of constant depth in the image, the visual angle according to which the pedestrian is seen is materialized by a cylinder of height h and radius r for a camera 30 which is positioned at a height H.

It appears that this angle varies according to the position of the pedestrian in the image. The considered band of the image is considered extreme values that are not too far from each other. For this, it is fixed a maximum angular variation in the band corresponding to a pixel. By way of illustration, two bands B1 and B2 are represented in FIG. 2, the two bands B1 and B2 being relative to the predetermined spatial zone ZS.

At the end of the correction step S106a, the defects related to the particular shooting of the camera 30 were taken into account in the analysis of the synthetic images.

At the segmentation step S106b of the crowd, the pixels belonging to the crowd are determined throughout the image. For this, in the learning base, it is used ground truth data.

In the extraction step S106c, the characteristics are extracted.

The characteristics are related to a particular pixel p, the pixel p is a segmented pixel at the segmentation step S106b. This ensures that the pixel p is connected to the presence of a crowd.

The characteristic attached to a pixel p is calculated on a sliding window W ^P _TO of radius r _w centered on the pixel p.

Extracted features are of two types, outline features and texture characteristics.

By way of example, the contour characteristics are obtained using a Canny contour detector, more often referred to as "Canny edge detector".

The contour characteristics are, for example, the orientation of the contours, or their length.

Alternatively, the extraction step S106c of the contour features is implemented using a Minkowski fractal dimension. Such a technique is also noted according to the acronym MFD for the English Minkowski Fatal Dimension.

The extraction step S106c of the texture characteristics is, for example, implemented using grayscale dependency matrices. Such matrices are usually referred to by the acronym GLDM for "Gray Level Dependency Matrix".

The extraction step S106c also comprises a step of presenting the extracted characteristics in the form of a vector having a plurality of dimensions by differentiating the radius values r _w .

More precisely, it is considered a reference value for the radius r _w corresponding to one third of the half of the apparent average height in the image of a human being as described above for the coefficient σ. Such reference values denoted r _ref are used to obtain six different values of radius r _w which are, in the example of FIG. r _ref , 1/2. r _ref , r _ref , 3/2. r _ref , 2. _ref and 3. _ref .

Therefore, according to an implementation by the applicant, for each pixel of the image, there is obtained a characteristic vector of dimensions 1 ^* 144.

In the learning step S106d, it is used, according to the illustrated example, a vector vector regression. Such a regression is also noted SVR, acronym of the English Support Vector Regression. Such a regression binds the feature vector to the values taken by the spatial density in the density function F.

The performances of the SVR technique depend deeply on a good tuning of parameters that are chosen both in the kernel and the technique itself.

In the illustrated example, the kernel used is a Gaussian RBF kernel, RBF being the acronym for "Radial Basis Function", which means "radial-based function". For the rest, the kernel is noted k kernel.

To characterize such a kernel, three parameters must be characterized: a kernel parameter γ, a penalty parameter of the error term C and a loss function parameter ε.

The kernel parameter γ is chosen so that the k kernel satisfies the following equation: k (x _p , x _q ) = exp (-γ. || x _p - x _q || ² )

Where x _p is x _q are the p-th and q-th feature vectors.

The optimization of the choice of the three parameters makes it possible to obtain good performances. A common method for obtaining good parameters is split cross-validation with the learning set.

Since cross-validation is expensive in terms of time, the following constraints have been used for the penalty parameter of C and the parameter of the loss function ε.

Jlogn

not

OR :

- y is the average value of the field truth values,

_{where y} is the standard deviation of the field truth values,

- o _b is the standard deviation of the noise of the field truth values,

n is the number of training data, and - τ is a constant value.

It was chosen by the applicant to set the constant value r to three and for the core parameter γ to set the parameter to the value of the median of the list of all the values of the square of the norm between x _p and x _q for all possible p and q.

At the end of the learning step S106d, the law is obtained making it possible to estimate the spatial density in the predetermined zone.

In the particular case presented, the law is a regression model since the learning technique used is a SVR technique.

The second phase P2 of enumeration in the zone comprises three stages; an acquisition step S202, a obtaining step S204 and an estimation step S206.

At the acquisition step S202, an image is acquired by the detector 30.

The image has at least the predetermined spatial area ZS.

In the obtaining step S204, the characteristics determined at the extraction step S106c are obtained for the image supplied at the acquisition step S202.

The obtaining step S204 comprises a segmentation step S204a and an extraction step S204b.

The segmentation step S204a is performed by using a crowd detection algorithm in the image.

The segmentation operation extracts characteristics that are relative to the crowd and not features relating to places where there is no crowd. Improved segmentation increases the overall performance of the enumeration process.

The extraction step S204b is similar to the extraction step S106c described for the first phase P1.

Since the extraction is carried out only for the pixels for which a crowd has been detected, at the end of the obtaining step S204, the characteristic vectors are obtained at each pixel of the image for which a crowd has been detected.

At the estimation step S206, the spatial density of objects in the predetermined area is estimated. For this, it is applied the law obtained at the end of the first phase P1 to the acquired image. This link is shown by the arrow 50 in FIG.

More precisely, the regression model obtained at the end of the first phase P1 is applied to obtain a spatial density value. Otherwise formulated, the model is applied to each characteristic vector to deduce an estimate of the value taken by the function F at the corresponding pixel p. Moreover, since the estimation is only performed in the segmented zones and the function F is the result of a Gaussian convolution, the function F may in some cases also have significant values outside the zone. segmented.

To estimate the values taken by the function F outside the segmented area, the obtained spatial density values are extrapolated using image convolution using a normalized kernel K such that:

(/ - u ₀ ) ² + (k ₂ - v ₀ ) ²

V (/, / ^), ^ _! , / ^) = exp (-) with

the constant u ₀ defined by the equation u ₀ = u _c - β. (k _t - _c );

the constant v ₀ is defined by the equation v ₀ = v _c - β. (k ₂ - v _c ), and

the constant β is defined by the equation β = a. , °

^ (k ₁ -u _c ) ² Hk ₂ -Vc) ²

- u _c , v _c are the coordinates of the center of the core, these coordinates being adjustable by the processing module 10.

The parameter a is determined using the images obtained by the learning parameter.

Parameter a is, for example, found by a dichotomy technique.

At the end of the estimation step S206, an estimate of the spatial density of the number of pedestrians in the predetermined spatial zone ZS is obtained.

Optionally, the second phase P2 comprises a deduction step during which the number of pedestrians is deduced from the estimated spatial density at the estimation step.

For example, the number of pedestrians in the predetermined area is obtained by integration over the entire predetermined area. The method allows both to obtain analysis elements and behaves as a self-learning system.

Compared to the state of the art method, the proposed method uses a learning method implying, in fact, a database easier to obtain.

In fact, in the methods of the state of the art, the learning method is done using real data, which implies a manual annotation by an operator of the data to obtain a learning base.

In addition, the learning base of the proposed method has a better quality than the learning base of the method of the state of the art. Indeed, the variability and representativeness of the training data are questionable for the processes of the state of the art. In particular, in practice, in the basics of the state of the technique, extreme situations ie a spatial density greater than 50 people in the area are rarely present in the actual data because the situation is too rarely in practice. In addition, the annotation of the operator is actually unreliable especially when the crowd is very dense or the resolution is limited, these problems can be cumulated. Conversely, the base is generated to be representative of the future use of the detector 30, that is to say that the different images generated can scan very different densities of space ranging from low to very dense.

In addition, the method has the advantage of being robust to the change of position and / or orientation of the detector. Indeed, the shooting parameters are respected. The training data is virtually created with the same camera 30 that will be used next. For this, it is provided data to the camera 30 as the intrinsic characteristics of P _in tnnsèque focal or image or extrinsic size of the P _position and P _or orientation ientation- With the parameters given by the camera 30 , the learning database is automatically constructed in a faithful and statistically representative way.

The robustness is also ensured by the fact that a change made on the camera 30 only supposes to implement the first phase P1 without acquiring field images which would be indispensable in a method of the state of the art.

The method is also advantageous in that the ground truth is immediately available since the annotation is automatic and any camera change can be immediately taken into account via the learning of a regression model not assuming the provision of a new learning game.

The method applies in particular to the supervision of crowds by means of video surveillance in a security or transport perspective. In such an application, the method makes it possible to estimate the spatial density of a crowd and to count the number of pedestrians in the crowd.

Frequently, in such a situation, the acquisition step is done continuously so that it can be considered that the second phase P2 is performed iteratively on a continuous video stream.

The proposed method is related to computer-assisted vision, the computer implementing the learning. The method can therefore be implemented using any computer or any other device. Multiple systems can be used with programs implementing the above method but it is also conceivable to use devices dedicated to the implementation of the previous method, these may fit into devices that measure the data provided. In particular, the method can be implemented by a monitoring device capable of counting the objects in a predetermined spatial zone ZS comprising a detector 30 for acquiring at least one image, the image comprising the predetermined spatial zone ZS and a module of processing 10. The processing module 10 is adapted to count objects in the predetermined spatial zone ZS and is able to implement the steps of choice, supply S102, generation S104, determination S106, obtaining S204 and estimate S206. The processing module 10 is also adapted to receive parameters relating to the detector 30, the received parameters comprising at least one parameter being relative to the focal length of the detector 30, a parameter being relative to the altitude of the detector 30, a parameter being relative at the inclination of the detector 30 relative to the predetermined spatial area ZS.

In addition, the proposed embodiments are not related to a particular programming language. Incidentally, this implies that multiple programming languages can be used to implement the previously detailed method.

The method and embodiments described below are capable of being combined with each other, totally or partially, to give rise to other embodiments of the invention.

Claims

1. - Method for counting objects in a predetermined spatial area (ZS), the method comprising:

a first unsupervised learning phase comprising the steps of:

selecting a plurality of distinct spatial densities of objects in the predetermined spatial area (ZS),

- providing parameters relating to a detector (30) suitable for acquiring at least one image, the image comprising the predetermined spatial zone (ZS), the parameters provided comprising at least one parameter relating to the focal length of the detector (30), a detector altitude parameter (30) and a parameter relating to the inclination of the detector (30) relative to the predetermined spatial area (ZS),

generating a database comprising a plurality of images for each selected spatial density, the generation step using the parameters provided in the supplying step, and

determining a law making it possible to estimate the spatial density in the predetermined zone by using the generated learning base,

a second enumeration phase in the predetermined spatial zone (ZS) comprising the steps of:

acquiring an image comprising the predetermined spatial zone (ZS) by the detector (30), and

estimating the spatial density of the object in the predetermined spatial zone (ZS) by applying the law to the acquired image.

2. - The method of claim 1, wherein the generating step comprises a step of correcting the perspective defect related to the detector (30).

The method of claim 1 or 2, wherein the method includes a step of segmenting an image to determine the pixels containing an object.

4. - Method according to any one of claims 1 to 3, wherein the determining step is implemented by learning, the learning comprising a feature extraction, the extracted features are also used in the step of estimate.

The method of any one of claims 1 to 4, wherein the determining step is performed using a carrier vector regression.

The method of any one of claims 1 to 5, wherein in the providing step, a parameter relating to the detector (30) is an intrinsic characteristic

(Pintrinsec) to the detector (30).

7. - Method according to any one of claims 1 to 6, wherein in the supplying step, there is provided a parameter relating to the position (P _{po ition} ) of the detector (30) and / or a parameter of orientation (P _or ientation) of the detector (30).

8. - Computer program product comprising software instructions, the software instructions implementing a method according to any one of claims 1 to 7, when the software instructions are executed by a computer.

9. - Information carrier on which is stored a computer program product according to claim 8.

10. - Monitoring device for counting the objects in a predetermined spatial area (ZS) comprising:

a detector (30) capable of acquiring at least one image, the image comprising the predetermined spatial zone (ZS), the detector (30) having parameters, a parameter being relative to the focal length of the detector (30), a parameter being relative to the altitude of the detector (30) and a parameter being relative to the inclination of the detector (30) relative to the predetermined spatial area (ZS), and

a processing module (10) adapted to count the objects in the predetermined spatial zone (ZS),

the treatment module (10) being able to:

choosing a plurality of distinct spatial density in the predetermined spatial area (ZS),

receiving parameters relating to the detector (30), the parameters received comprising at least the parameter relating to the focal length of the detector (30), the parameter relating to the altitude of the detector (30), the parameter relating to the inclination of the detector (30) with respect to the predetermined spatial area (ZS), generating a plurality of images for each selected spatial density using the parameters received,

determining a law making it possible to estimate the spatial density in the predetermined spatial zone (ZS) by using as basis of learning the generated learning base, and