EP0577491B1 - Process and device for monitoring a three-dimensional scene using imagery sensors - Google Patents

Description

The present invention relates to surveillance a lit, three-dimensional scene with structure known or determinable geometric, allowing in particular to detect anomalies larger than a value determined (representing for example intrusions) or, in an advanced embodiment, to locate a mobile moving inside the scene.

We already know many monitoring methods (see GB-A2 150 724 and R.J. Blissett: "Retrieving 3D information from video for robot control and surveillance ", in Electronics and Communication Engineering Journal, pages 155-164, August 1990, Vol.2 n ° 4) of a three-dimensional illuminated scene, the illumination being due to a natural source, such as the sun, or a artificial source of electromagnetic radiation or ultrasound. Many of these processes compare between the image of the scene at a given time and the image previously formed. These processes have in particular the disadvantage of poorly discriminating local modifications due to intrusion from those due to local modification of illumination. Generally, these methods and devices most have insufficient selectivity for allow to differentiate, with a high probability, anomalies characterized by their size, and in particular their protruding from the original scene, so from differentiate, for example, a rodent and a human being.

The present invention aims to provide a method and a scene monitoring or control device, meeting the requirements better than those previously known of practice, particularly in that they provide high selectivity by using only means simple hardware.

• on constitue un modèle géométrique tridimensionnel (dit 3D) de référence de ladite scène associant, à chaque point de la scène, des points homologues dans le champ d'au moins deux capteurs placés à distance l'un de l'autre et sensibles à un paramètre radiométrique (tel que la brillance) de la scène ;
• on synthétise périodiquement la scène réelle sur le modèle géométrique tridimensionnel, à partir de chacune des représentations qui en sont fournies par les capteurs ; et, à partir de là,
• on compare les scènes synthétisées entre elles pour faire apparaítre les écarts dus à une modification du modèle 3D qui ont une taille supérieure à une valeur déterminée.

To this end, the invention proposes in particular a method for monitoring a lit scene in three dimensions, according to which:

• a three-dimensional (so-called 3D) geometric model of said scene is constituted, associating, at each point of the scene, homologous points in the field of at least two sensors placed at a distance from each other and sensitive to a radiometric parameter (such as brightness) of the scene;
• the real scene is periodically synthesized on the three-dimensional geometric model, from each of the representations provided by the sensors; and from there
• we compare the synthesized scenes with each other to reveal the differences due to a modification of the 3D model which have a size greater than a determined value.

We see that the reference model is purely geometric and only intervenes to establish the correspondence between homologous points in the fields of different sensors. The comparison can be made between images taken at identical times, so in lighting conditions which are the same. Given that the comparison is made between synthesized scenes, a modification of the external conditions modifying the radiometric values (brightness or color) of a point will be has no effect on the comparison, provided that the sensors have similar response characteristics.

Incidentally, it should be noted that the term "scene" does not should not be interpreted as limited to the case of a zone continuous, corresponding to the field overlap sensors. The stage can be made up of portions discreet in the fields, only considered important and monitor.

The comparison of the synthesized scenes can be performed by determining the coefficients of correlation between the two synthesized scenes, on determined neighborhoods of the points of the scene, and by comparison of these coefficients to a threshold which will be obviously always less than unity.

The 3D geometric model of the scene can be constituted from the images of the scene, in a state of reference, provided by the sensors, stored either pixel per pixel or, more generally, using elements geometric such as facets.

The geometric model can also be established directly in mathematical form from a field knowledge provided by observations earlier or using maps and models of two-dimensional geometric representation by sensors can be derived from knowledge the position of the sensors relative to the scene.

In most cases, we will have to give sensors a mutual spacing less than the distance between the sensors and the points of the scene which are the most close: the relationship between the spacing of the sensors and the distance will be a paramount parameter to select the minimum projection that an object must have in relation to the original scene to be detected.

The detection operations defined above can be repeated, for example at the rate of taking view of the sensors, which will usually be cameras. So we can follow the evolution of an object in the scene.

The invention also provides a device for monitoring to implement the above process defined. This device includes several sensors sensitive to a radiometric parameter, in lengths comparable waveforms, including storage means a purely geometric three-dimensional reference model of a scene to watch, associating, at each point of the scene, homologous points in the sensor field; of means allowing, from the geometric model and signals representative of the radiometric value of each image pixel provided by the sensors, to synthesize the real scene by projection on the geometric model three-dimensional; and neighborhood comparison means of determined dimensions of the synthesized scenes, allowing to determine the gaps between them.

The invention is susceptible of numerous modes of different implementation, allowing to adapt it as well in case we can have or get a geometric model three-dimensional as a digital terrain model that in case it is necessary to determine a transformation establishing, at least approximately, a correspondence between each point of the field of one of the sensors with the homologous point of the field of the other sensor.

• la figure 1 montre schématiquement la disposition de deux capteurs par rapport à une scène et l'effet d'une anomalie en relief sur la scène, dans le cas d'un dispositif de surveillance que l'on peut qualifier de "fond de terrain" ;
• la figure 2 est un schéma destiné à faire apparaítre l'influence de différents paramètres sur la détection ;
• la figure 3 est un organigramme d'un mode de réalisation du procédé ;
• la figure 4 est un schéma de principe montrant une constitution possible d'un dispositif ;
• les figures 5A et 5B montrent des dispositions particulières, avantageuses du point de vue du traitement, des images fournies par les caméras constituant les capteurs d'un dispositif ;
• la figure 6 montre une architecture possible de la chaíne de traitement d'un dispositif.

The invention will be better understood on reading the following description of particular embodiments, given by way of nonlimiting example. The description refers to the accompanying drawings, in which:

• FIG. 1 schematically shows the arrangement of two sensors in relation to a scene and the effect of an anomaly in relief on the scene, in the case of a monitoring device which can be described as "background";
• Figure 2 is a diagram for showing the influence of different parameters on the detection;
• Figure 3 is a flow diagram of an embodiment of the method;
• Figure 4 is a block diagram showing a possible constitution of a device;
• FIGS. 5A and 5B show particular arrangements, advantageous from the point of view of processing, of the images supplied by the cameras constituting the sensors of a device;
• Figure 6 shows a possible architecture of the processing chain of a device.

Before describing in detail the steps of a method according to the invention for intrusion detection, the general approach implemented in accordance with the invention will be defined, in the particular case of a scene consisting of a terrain whose it is possible to represent the structure by a geometric model in the form of a digital terrain model (abbreviated DEM), giving the coordinates of each point of the terrain in a fixed axis system x , y and z .

FIG. 1 shows a device having only two sensors 10 and 12, constituted for example by sensitive cameras in the visible range, which have angular fields which overlap in the scene to be observed. In the image supplied by a sensor of order i (i being equal to 1 or 2 in the case considered) a point of the digital terrain model having coordinates (x, y, z) in the referential axis system chosen will correspond to a point of coordinates (αi, βi) and of radiometry Si in a two-dimensional frame of reference which is associated this time with the respective sensor. It is possible to determine, by calculation or prior measurements, the transformations making it possible to pass coordinates and radiometry in a frame of reference linked to the scene to coordinates and radiometry in a frame of reference linked to each sensor.

Knowing the 3D geometric model of the scene at observe and distribute the given radiometric values by one of the sensors, we can build a synthesis scene associated with this sensor, reassigning to each known point in the scene, i.e. at each coordinate point (x, y, z), the radiometric value supplied by the sensor.

A scene is thus constituted for each sensor synthesis by associating the real geometry, provided by the DEM, and the radiometric value, is directly provided by the sensor, is subjected to a preliminary treatment, by filtering example.

If the 3D geometry of the real scene has not undergone modification in a specific area, the scenes synthesized will remain almost identical to each other for this area, even if there has been an evolution in radiometry real. Even if, for example due to a different answer of the two sensors and lighting variations, the scenes synthesized do not remain strictly identical, the synthesized scenes will in any case remain closely correlated, in the absence of modification of the geometry of the ground.

If, on the other hand, the observed scene has undergone geometric modifications compared to the known model, by following an intrusion, the corresponding radiometric value at a point of the disturbance, seen from one of the sensors, will projected during the synthesis on a geometric position of the where it did not come from, since the device only knows the geometric model not having this disturbance. It will be the same for the value radiometric supplied, for the same point of the disturbance, by the other sensor. Consequently the synthesis scenes developed by two sensors, each based on information radiometric he receives and the geometric model of reference, will differ. The gaps between the two scenes of synthesis will depend on the height and volume of the disturbance.

This influence of the height of the disturbance appears immediately if you consider the particular case a disturbance having only a very small thickness: in this case there is only modification of the values radiometric measurements in a particular area of the real scene. This scene will remain geometrically identical to the model of reference. The disturbance, like a disturbance that would due to a local change in lighting (passage of a drop shadow for example) will not be detected as a intrusion.

The device can therefore be robust to variations in lighting conditions and be insensitive to passage of small objects on the ground.

We will see below that the size, and in particular the height, minimum detected depends in particular on the resolution sensors and the distance between them.

It can also be seen in FIG. 1 that the anomaly 14 shown in dashes results in a modification of the image provided by the sensor 10 over the width l , which can be translated as the addition of a shadow on the picture. The same anomaly results for the image provided by the sensor 12, by a modification on the width l : the synthesis scenes will therefore no longer coincide on all of the "shadows". This remains true whatever the nature of the lighting, whether optical, electromagnetic at high frequency or even ultrasonic.

The detection sensitivity of the device depends on the resolution at the level where intrusions are likely to occur. If, for example, we denote by D the interval between two sensors, assumed to be identical, and by E _i = Pi Pi + 1 the resolution at ground level, for point i and sensor 12, an anomaly will be detected on the sole condition that it extends beyond the angular resolution domains of the two sensors, associated with the same location of the terrain model. Consequently, the anomaly indicated in 14 ₁ will not be detected, whereas the anomalies 14 ₂ and 14 ₃ may be. Consequently, the conditions to be met for detecting an anomaly can be determined as soon as it exceeds a determined height above the ground.

If in particular the terrain is far from the sensors and is not too disturbed, we can consider that the minimum height d with respect to the terrain that an obstacle must have to be detected is related to the other parameters by: D / (Ld) = (L.pix) / d where L is the average distance of the sensors to the ground, D is the distance between the sensors, and pix denotes the angular resolution corresponding to the resolution on the ground E _i .

In the case of a device that can be described as "ground floor" intended to detect intruders on the ground, while avoiding false alarms due to small animals, L will always be much greater than d . We can then write: L 2 = Dd / pix.

As an example, we can adopt, to detect objects larger than 10 centimeters, d = 20 cm.

The result can then be achieved with:
pix = 10 ^-3 rad
L = 20 m
D = 2 m.

Whatever solution is adopted, we see that uses knowledge of the geometric structure of the scene to give the detection a comparison character geometric, which increases performance. The knowledge that a priori eliminates the influence of known elements and favors the detection of modifications geometric due to intrusions.

The device can be supplemented by sensors other natures, or implement multiple systems associates working in different wavelengths. In addition, it can implement, in addition to detection using a comparison between radiometric values supplied by two or more sensors (e.g. infrared), a time detection using a comparison between successive images.

Finally, once an object is detected, it is possible to make it disappear by incorporating its characteristics in the geometric model of the observed area, either by a new shot and an analysis of the new views, i.e. purely mathematically.

The process, as just described so far involves knowledge of the 3D geometric model of the scene observed to develop the synthesis scenes.

But it suffices, to implement the method, to have a geometric correspondence between the pixels images of the various sensors. This correspondence geometric is established for the reference scene in order to then detect disturbances. Any disruption geometry of the reference scene causes a difference radiometric between the points associated by the correspondence in the reference scene. As we will see later, this correspondence can be established from the identification in the image of landmarks in the fields of the various sensors. Finally, a model can be established by equating the terrain with a mosaic of facets each having a simple geometric shape.

Under these conditions, from the approach set out above, various detection methods are likely to be implemented. We will now mention two solutions, which differ in the method of implementation geometric correspondence of the model and the image synthesized.

The first method assumes that a digital model of the terrain has been produced and stored. This digital model can be created from aerial photographs, maps, or even measurements in the field. From the knowledge of this mathematical model, we can determine by calculation, for a given position of each of the sensors, at least two for each scene, the transformations which correspond, to each point of coordinates (x, y, z) of the digital model, the coordinates αi, βi of the image of this point in the coordinate system associated with the sensor of order i .

In other words, we can write, by designating by T the determinable transforms: αi = Tα i (X Y Z) βi = Tβ i (X Y Z)

The radiometric signal S (αi, βi) supplied at each point of the image of the sensor i is proportional to the flux re-emitted by the surface element associated with the points (x, y, z) of the model, the coefficient of proportionality being substantially constant over a neighborhood of a few pixels.

The detection scheme, in the case of two sensors 10 and 12 constituted by two cameras, is then that shown in FIG. 3. At a point (x, y, z) diffusing a flow S , it is possible to associate the image points (α1 , β1) of camera 1 and (α2, β2) of camera 2. We can also deduce the signals S1 and S2 from this.

• des coordonnées (x,y,z) qui ne varient pas dans le temps, puisqu'elles correspondent à la géométrie réelle de la scène,
• des radiométries qui fluctuent suivant les conditions d'éclairement.

We can thus create two synthetic objects each corresponding to one of the cameras, consisting of points each having:

• coordinates (x, y, z) which do not vary over time, since they correspond to the real geometry of the scene,
• radiometry which fluctuates according to the lighting conditions.

The detection operation is carried out by comparison signals S1 and S2 associated with the same point (x, y, z) and this for each of the points (x, y, z).

V(p) désigne un voisinage des points P respectifs, en général sur 3x3 ou 5x5 pixels dans le cas d'un réseau carré ;

p' désigne un point courant de V(p), de coordonnées (x',y'z').

This comparison can be made by various mathematical operations. One solution giving good results consists in calculating the coefficient C (p) of local normalized intercorrelation between the two synthesis models, in the form: C (p) = Σ S 1 (T α 1 (P '), T β 1 (P ') S 2 (T α 2 (p '), T β 2 (P ')) p' ε v (p) [Σ S 1 (T α 1 (P '), T β 1 (P ')) 2 x Σ S 2 (T α 2 (P '), T β 2 (P ')) 2 ] 1/2 p ' ε v / (p) or

V (p) designates a neighborhood of the respective points P, in general on 3 × 3 or 5 × 5 pixels in the case of a square array;

p 'denotes a current point of V (p), with coordinates (x', y'z ').

The normalized cross-correlation coefficient is then equal to 1 if, for a given neighborhood V (p), we have: S 1 (T 1 α (p '), T β 1 (p ')) = λ S 2 (T α 2 (p '), T β 2 (p '))

This equality is respected if the scene observed is not disturbed.

This method makes it possible to get rid of consequences of differences in gains between the sensors and effects due to angular conditions distinct from shooting, having an influence on the backscatter, towards the sensors.

Insofar as the observed scene is perfectly diffusing, all it takes is a pixel comparison to pixel of the two images.

In the event of an intrusion, the synthesis model is reconstructed with a radiometry which is no longer that associated with the geometry of the observed area. We associate in effect, at a point of coordinates (x, y, z) hidden by the intruder, a radiometry which did not come from the point of the scene of reference. And this is true for all the points masked by the perturbation.

Detection is determined by issuing an alarm signal when the correlation coefficient is less than a threshold s , which is obviously always chosen to be less than 1.

Such a detection technique has the advantage to be insensitive to variations in dynamics and in particular of illumination, which constitutes a factor of robustness important and allows omitting any radiometric calibration.

A second approach should be used when we do not initially know a numerical model of field: it allows to return to the initial conditions of the where a DTM is available.

• {α_k1, β_k1} les coordonnées des images du point d'ordre k dans le repère lié à la caméra 1,
• {α_k2, β_k2} les coordonnées des images dans le repère lié à la caméra 2.

One solution is to use remarkable points in the scene, easily spotted in each image and whose coordinates (x, y, z) are known in the same geographic frame, constituting bitter points. We will designate by {x _k , y _k , z _k } the coordinates of the bitter point of order k and by:

• {α _k1 , β _k1 } the coordinates of the images of the point of order k in the coordinate system linked to the camera 1,
• {α _k2 , β _k2 } the coordinates of the images in the coordinate system linked to camera 2.

We then look for the transformation allowing to pass from coordinates in the real scene to coordinates in the image: - α i = T αi (X Y Z) - β i = T βi (X Y Z).

Depending on the complexity of the scene, we will have to adopt equations of more or less complex forms for the transformations T. The number of bitter points necessary to give a sufficiently accurate representation will increase with the complexity of the transformation. When the scene is not too tormented, it is possible to adopt transformations T _α and T _β of bilinear form: Tα (x, y, z) = (α000) + (α100) x + (α010) y + (α001) z + (α110) xy + (α101) xz + (α011) yz Tβ (x, y, z) = (β000) + (β100) x + (β010) y + (β001) z + (β110) xy + (β101) xz + (β011) yz

The coefficients α000, ..., β011 are calculated by writing, for all the bitter points, the equations: - α k i = T α i (x k , y k , z k ) - β k i = T β i (x k , y k , z k )

We thus obtain systems of linear equations which can be solved without difficulty: in the above case seven coefficients are to be determined, which requires the minus seven points of bitter. A larger number may be used, so as to retain coefficients representing a optimum.

Once the transformations T _α and T _β known, we find ourselves in the previous case, since we know the equivalent of a DTM.

Another solution is to use a grid of transformation, projected onto the real scene and which mark the representation in each image. This grid can in particular be constituted by a wrought, obtained by example by scanning with a laser and location, each laser pulse, from the point of the image representing impact on the ground.

We will now describe, with reference to the Figure 4, the basic principle of a surveillance, intended for automatic intrusion detection in known areas. Figure 4 shows two sensors, constituted by cameras 10 and 12 whose fields have an overlap zone 16 of which a part or the all constitutes the monitored scene. Lighting can be natural; at night or to improve the sensitivity of short distance detection, natural lighting can be replaced or supplemented by an artificial, visible or infrared. In the case of person detection, it suffices with a ground resolution of 10x10 cm and a cadence of two detections per seconds, which corresponds to a displacement of 1.5 m for one person at 10 km / h. The sensors can be CCD infrared cameras with a matrix of 600x700 pixels. The two cameras 10 and 12 are synchronized, by example by controlling camera 12 from the signal from camera synchronization 10.

• deux cartes d'acquisition 20 et 22 permettant de récupérer la partie utile des images des caméras,
• une carte processeur de signal 24,
• une carte de mémoire vive 26, et, éventuellement,
• une carte de communication 28 permettant de connecter l'unité de traitement à un réseau téléphonique.

The video signals coming from the cameras are applied to a processing unit 18 which can be viewed as comprising:

• two acquisition cards 20 and 22 making it possible to recover the useful part of the images from the cameras,
• a signal processor card 24,
• a RAM card 26, and possibly
• a communication card 28 making it possible to connect the processing unit to a telephone network.

Although we can arrange cameras 10 and 12 in any relative orientation it is however advantageous, to reduce the complexity of processing and the memory capacity required, to respect constraints which in practice are not annoying and some of which will now be on display.

First of all, it will be recalled that the corresponding pixels of two images supplied from the same scene by two cameras are on two straight lines, called epipolar, each passing through the optical center O ₁ and 0 ₂ of the respective image: in each image , the pixel corresponding to the same point P of the scene is in the image - constituted by an epipolar line - of the plane 0 ₁ PO ₂ . The epipolars have the property that any pixel of the epipolar in one image has a homologous pixel in the corresponding epipolar of the other image if the reference marks of the cameras remain unchanged.

An interesting configuration is to place the cameras so that their marks (line and frame directions) are parallel and that the points 0 ₁ and 0 ₂ are located on the same straight line parallel to the line direction. Two epipoles such as 30 ₁ and 30 ₂ are then straight lines parallel to the line direction (FIG. 5A). The treatments can then be carried out line by line according to the epipolar: it suffices to memorize in the card 26 the lines corresponding to the epipolar of the zones observed and a number of lines above and below corresponding to the extent of the chosen neighborhood for correlation calculation.

When it is not possible or inconvenient to adopt the arrangement of FIG. 5A, that of FIG. 5B is still advantageous. The epipolar 32 ₁ and 32 ₂ are then inclined lines, but making the same angle with respect to the lines of the two images.

It is then necessary to carry out a treatment additional, compared to Figure 5A. Or else selects the homologous points of the epipolar and parallel to the epipoles belonging to the chosen neighborhood, or the memory size 26 is increased.

FIG. 6 shows a possible constitution of the processing part of a device in the case of a terrain model using a memorized grid. It will be assumed that the optical centers of the cameras 10 and 12 are arranged in the vicinity of the same straight line parallel to the direction of scanning in the line direction. In this case the memory 26 (FIG. 4) can comprise two memory spaces 26 ₁ and 26 ₂ corresponding to a packet of a few lines of pixels only, centered on the line of order i whose pixels are neighboring. The processor can be viewed as having two processing channels 24 ₁ and 24 ₂ performing the projections and receiving, on respective inputs 34 ₁ and 34 ₂ , the geometric models in digital form.

The processor also includes a unit 36 of comparison of projections, usually by determination an intercorrelation factor on each neighborhood and comparison to a threshold.

The output of the unit 36 is applied to a block 38 for transferring the images to a recording device and possibly a display device 40. The processor 36 also has an output 40 controlling the passage of a packet of lines centered on line i has a packet centered on line i + 1 when the comparison is complete. In other words, the processor drives the storage of line packets asynchronously.

The architecture shown in Figure 6 allows limit the amount of RAM to write and read fast.

The invention is susceptible of numerous variants of realization and many applications other than those that have been specifically mentioned.

For example, the device can be used to locate a robot inside the scene, the robot constituting an anomaly.

In this case, it may be advantageous to place the sensors on a displacement device (bridge carriage rolling for example) fitted with measuring means allowing permanently have the coordinates of the sensors by report to the field. It is thus possible, from a main memory storing a digital terrain model corresponding to the entire robot evolution domain, to constantly update the processor memory of processing to match the location of the sensors. The device according to the invention has the advantage to be free from most system faults visualization mounted on the robots themselves, including the limitation of the distance field of vision and the impossibility to observe the area of movement of the robot behind obstacles.

Claims (11)

1. A method for the surveillance of a three-dimensional scene having a known or determinable geometrical structure, making it possible to detect the appearance of an anomaly in relief and including the following stages:

   a three-dimensional geometrical reference model of the said scene is made up and memorised, in a reference state, associating with each point of the scene homologous points in the field of at least two sensors placed at a distance from each other and sensitive to a radiometric parameter (such as the brilliance) of the scene;

   periodically, the real scene is synthesised over the three-dimensional geometrical model on a basis of each of the representations thereof which are furnished by the sensors in order to obtain as many synthesised scenes as there are sensors, and

   points of closeness belonging to the synthesised scenes are compared with one another in order to show up divergencies due to a modification of the three-dimensional geometrical model, only those divergencies which correspond to a minimal height anomaly in excess of a specific value being retained.
2. A method according to claim 1, characterised in that the synthesised scenes are compared by determining coefficients of intercorrelation between the two synthesised scenes over specific closenesses of the points of the scene and in that the coefficients obtained are compared with a threshold.
3. A method according to claim 1 or 2, characterised in that the geometrical model of the scene is made up on the basis of images of the scene, in a reference state provided by the sensors and memorised pixel by pixel.
4. A method according to claim 1 or 2, characterised in that the geometrical model is established in mathematical form on the basis of knowledge of the terrain supplied by prior observations or with the help of maps.
5. A method according to claim 1 or 2, characterised in that the geometrical model of the scene is made up on the basis of landmark points of known co-ordinates (x, y, z) and an equation of predetermined form
6. A method according to claim 1 or 2, characterised in that the geometrical model is made up on the basis of a currying grid produced on the scene by scanning using a laser and recorded in the form of the location for each sensor, at each laser pulse, in respect of that point of the image which represents the impact on the terrain.
7. A method according to any one of the preceding claims, characterised in that, in addition, successive images delivered by the sensors are compared in order to carry out a time-related detection.
8. A method according to any one of the preceding claims, characterised in that the sensors are spaced apart mutually by an amount less than the distance between the sensors and the closest points of the scene.
9. A device for surveillance of a three-dimensional illuminated scene having a known or determinable geometrical structure, making it possible to detect relief anomalies over the scene, comprising at least two sensors (10, 12) sensitive to a radiometric parameter, in comparable wavelengths, and comprising likewise: means of memorising a three-dimensional purely geometrical reference model of the scene to be surveyed, associating with each point of the scene homologous points in the field of the sensors; means which make it possible to synthesise scenes in order to obtain as many synthesised scenes as there are sensors, each of the scenes being synthesised on the basis of the geometrical model and signals representing the radiometric value of each image pixel supplied by one of the sensors by projection on the three-dimensional geometrical model, and means of comparing points of closeness of predetermined size on the set of synthetic scenes on the basis of signals delivered by the various sensors, making it possible to determine the distances between the same locations in the two synthesised scenes.
10. A device according to claim 9, characterised in that the said sensors furthermore comprise sensors working in different wavelength ranges over the two said sensors.
11. A surveillance device according to claim 10 intended to locate a robot adapted for movement within the scene, characterised in that the said sensors are carried by means of displacement in respect of the scene and are associated with a memory for updating the geometrical model memorised as a function of the position of the sensors in respect of the scene.

Images