FR2985070A1 - System for detecting fall of e.g. person in retirement home, has programmed analysis unit to acquire images taken by sensors in synchronous manner, and to detect fall of person in scene from time sequence of pair of images of scene - Google Patents

Abstract

The system (1) has a digital camera device (2) comprising two visible-infrared array sensors (3, 4) differently viewing a same scene (10). A programmed analysis unit (5) acquires images taken by the sensors, where the images are taken synchronously and form a pair of images of the scene. The analysis unit detects the fall of a person in the scene from a time sequence of the pair of images of the scene. An infrared lighting device (7) ensures diffuse lighting in the scene. Independent claims are also included for the following: (1) a method for detecting fall of persons (2) a computer program product comprising a set of instructions for implementing a persons fall detection method.

Description

The present invention relates to the technical field of the detection of falls of persons by automatic image analysis. Here we mean by "fall" the action of falling to the ground in an involuntary way. In particular, various factors may be at the origin of this action such as age, state of health (walking or balance disorders, reduced visual acuity), behavioral factors (sedentary lifestyle, following a consumption of alcohol for example), and / or environmental factors (insufficient lighting, a crowded place of passage for example). On the other hand, the term "persons" refers to any subject at risk of falling, such as elderly people, sick people, isolated workers, young children, or robots. In particular, the fall is the major cause of death among the elderly. As solutions to this problem, there are different approaches in the state of the art using - techniques based on the use of kinematic / physiological drop sensors - to be worn permanently by the person generally in the form of medallions or bracelets - watches - programmed to signal the fall on the basis of the person's horizontal and / or vertical posture; or - video surveillance techniques based on the analysis of images captured by a plurality of cameras deployed in the environment to be monitored. The known methods and systems are imperfect, mainly because of their low reliability in terms of resistance to false alarms (false positives: the signaling of a fall then there is none) and the defects of detection of a fall (false negatives: not reporting a fall when there was one).

Indeed, the wide variety of possible falls of a person, some of which are similar to daily actions, is a source of error for most systems based on kinematic sensors worn by the person. As illustrative examples, the actions "sitting on the bed and lying down", "sitting on a chair from the standing position" and "bending or crouching to pick up an object on the floor" falling respectively from a seated position and ending in a lying position, to a fall ending in a seated position, and to a forward fall ending in a squatting position with recovery - may defect the performance of kinematic sensors of falls. In addition, the port of sensors can be seen as intrusive by the person, where he is obliged to wear at least one badge. With regard to video surveillance techniques for the detection of falls of persons, the presence of disturbing elements in the environment to be monitored (multi-source and variable lighting, darkness, low lighting, shadow of a person in motion in this environment, the presence of several people, the presence of pets, changes in this environment such as the movement / introduction of objects for example) are likely to limit the reliability of these techniques. In addition, the video surveillance-based fall detection systems known to date use a distributed system consisting of at least three cameras, making them complex and laborious to deploy. An object of the present invention is to overcome the aforementioned drawbacks. Another object of the present invention is to provide a real-time fall detection system capable of detecting the fall of a person in a reliable and relevant manner.

Another object of the present invention is to provide a fall detection method capable of tracking more than one subject in a supervised environment. Another object of the present invention is to provide a non-intrusive fall detection system (i.e. not requiring the port of the sensors). Another object of the present invention is to propose a system for detecting falls of persons insensitive to the disturbing elements mentioned above.

Another object of the present invention is to simplify fall detection systems by reducing the number of cameras required. Another object of the present invention is to provide a compact fall detection system. Another object of the present invention is to ensure the well-being of people at risk of falling. To this end, the invention relates, according to a first aspect, to a system for detecting falls of persons, comprising: a digital camera device provided with a first matrix sensor that is sensitive to infrared and visible; a second matrix sensor sensitive to visible infrared, the two sensors looking substantially differently the same scene; an analysis unit programmed to acquire, from said digital camera, an image taken by said first infrared visible sensitive matrix sensor and an image taken by said second infrared visible sensitive matrix sensor; these two images being taken synchronously and forming a pair of images of said scene; o detecting the fall of at least one person in said scene from a sequence in time of image pairs of said scene.

Advantageously, the first and second infrared visible sensitive matrix sensors are spatially calibrated relative to the ground of the scene to be monitored. This system further comprises an infrared lighting device arranged to provide diffuse lighting in the scene. According to one embodiment, the analysis unit is programmed to - detect moving objects in said scene from the images taken by said first infrared visible sensitive matrix sensor; tracking in two dimensions the objects detected in the images taken by said first infrared visible sensitive matrix sensor; skeletonizing, in at least one image taken by said first infrared visible sensitive matrix sensor, objects tracked in two dimensions; - discretize at least one skeleton; searching in the image taken by said second infrared visible sensitive matrix sensor, in a pair of images, at said at least one image taken by said first sensitive matrix sensor of the visible infrared of the dots counterparts at the discrete points of the discretized skeleton; calculating the 3D positions of the discrete points of said skeleton; - Compare with predefined values the rate of variation over time of the average heights of the discrete points. Advantageously, in order to detect the objects in motion, the analysis unit is arranged to use the variation in time of the contours from one image to the next in a sequence of successive images taken by said first sensitive visible matrix sensor. infrared. Advantageously, in order to follow in two dimensions the detected objects, the analysis unit is arranged to - smooth the detected contours; - identify related components in smoothed outlines; - to distinguish objects on the basis of the identified connected components; - to represent each object distinguished by a binary mask. Advantageously, in order to calculate the 3D positions of the discrete points 5 of said skeleton, the analysis unit uses the principle of epipolar geometry for a pair of images taken from two substantially different angles of view. According to a second aspect, the invention relates to a method for detecting falls of persons, comprising the steps of: synchronously taking a pair of digital images of a scene, an image in a first angle of view and an image at a second angle of view, the first and the second angle of view being substantially different; detecting the fall of at least one person in said scene from a time sequence of digital image pairs of said scene. According to a third aspect, the invention proposes a computer program product implemented on a memory medium that can be implemented in a computer processing unit and includes instructions for carrying out the method. summary above. Other features and advantages of the invention will appear more clearly and concretely on reading the following description of preferred embodiments, which is made with reference to the accompanying drawings in which: FIG. functional of a fall detection system according to one embodiment; Figure 2 schematically illustrates a non-limiting installation mode of the fall of persons detection system; Figure 3 schematically illustrates the 3D calibration step of the digital camera of the fall detection system; Figure 4 is a schematic diagram illustrating a possible mode of operation of the fall detection system; FIG. 5 schematically illustrates the effect of a shadow suppression step applied to an image of the environment to be monitored; Figures 6, 7, 9 and 10 schematically illustrate intermediate results of the method of detecting falls of persons. Figure 8 schematically illustrates the principle of epipolar geometry. Figure 11 shows a curve of the variation in the height of a moving person in the monitored environment. In Figure 1 is shown a system 1 arranged to detect and report in real time any person falling in the environment 10 under surveillance (a scene 10 under surveillance).

In particular, the environment 10 may be an indoor environment (a room, a corridor, a staircase, a hall, a subway train, a lift cabin, a subway station for example), or an outside environment (a parking, a train station, a bus station, the entrance / exit of a hospital for example).

The fall detection system 1 is, in fact, responsible for identifying and tracking the movement of any person present in the environment 10 in order to detect its possible fall and trigger, if necessary, an intervention protocol. For this, this system 1 for detecting falls of persons comprises - a digital camera 2 for taking pictures of the environment 10; an infrared lighting device 7; an analysis unit; and a communication module 6 enabling this system to interact with other communication systems. The camera device 2 comprises two matrix sensors 3-4 sensitive from the visible range to the infrared (two visible infrared visible cameras) juxtaposed so as to look substantially differently at the same scene 10 (or environment 10). ), as shown in Figure 1 or 2 (ie both cameras 3 and 4 have almost the same field of view). A matrix sensor 3-4 sensitive to the visible infrared allows in particular to capture, in the environment 10, the geometry and intensity of any observed physical phenomenon and its evolution over time. Advantageously, the two visible-to-infrared sensitive matrix sensors 3-4 assembled into a single device 2 offers the fall detection system 1 compactness and simplicity of deployment. The two visible-infrared matrix sensors 3-4 are arranged so that they look at the same scene synchronously, and are provided with wide-angle lenses so that their field of vision covers environment 10.

In an implementation, when the environment under surveillance is an indoor environment such as a room, the camera 2 is mounted at the top of a wall of the room, at the intersection with the ceiling, As shown in FIG. 2. In another implementation, when the monitored environment is an external environment, the camera 2 is placed at a predetermined height with respect to the environment 10 to be monitored. Furthermore, the infrared lighting device 7, preferably oriented towards the ceiling, makes it possible to have diffuse lighting and thus avoids the projection of shadow in the environment. Attenuation, see avoidance , the projection of shadows in the field of view of the device 2 digital shooting has the effect of allowing the detection of falls day and night with the same reliability. As a result, the fall detection system 1 is advantageously functional both during the day and at night without the need for any intervention on the part of the user (particularly the adjustment of the optics, the contrasts, the luminosities, for example). The optics is, in fact, used with quasi-maximum aperture and the amount of light that the sensitive matrix sensors 3-4 visible infrared receive for the day and night scenes is adjusted with the help of time. shutter. The analysis unit 5 comprises software components (ie, algorithms) and hardware components (memory spaces, a processor for example) whose conjugation makes it possible, on the basis of the images of the environment 10 acquired since the device 2 for shooting, to detect the presence of a person in the environment 10, to monitor the movements of this person, to detect its possible fall, and trigger, if necessary, the communication module 6. Indeed, the analysis unit 5 inputs images synchronously taken by the two cameras 3-4 (ie the two sensors 3-4 matrices susceptible to visible infrared) for the process in real time to detect a possible fall of a person followed in the environment 10 and subsequently emit an alarm via the communication module 6. It should be noted that prior to the startup in the environment 10 of the fall detection system 1, a preliminary 3D calibration step of the two cameras 3-4 is performed. This preliminary step makes it possible, in fact, to perform the spatial calibration of the matrix sensors 3-4 with respect to the ground of the monitored environment. In particular, this 3D calibration determines the positions and relative orientations of the two sensors 3-4 in the spatial reference linked to the environment 10. According to an embodiment illustrated in FIG. 3, the 3D calibration of the two cameras 3-4 is made using a pattern corresponding to a flat black and white checkerboard (or any other calibration chart) and whose square dimensions are known. This pattern is displayed synchronously on both sensors 3-4. A calibration algorithm detects the positions of the points of the test pattern and calculates the calibration parameters of the digital camera device 2. During the calibration process, the test pattern is placed on the ground to define the ground mark of the environment 10. The equation of the plane that minimizes the distance to all these points in the least squares sense is then formulated . Finally, a marker whose X and Y axes are on the plane corresponding to the ground of the environment 10, and whose Z axis is orthogonal to the plane is calculated. It should be noted that the mark (X, Y) is not necessarily horizontal, such as the case of monitoring a staircase for example.

Referring now to FIG. 4, the method of detecting falls of persons in the environment 10 - implementing the above-described components of the system 1 - comprises the following steps: - successive acquisition, by the unit 5 analysis, pairs of images taken synchronously, respectively by the two sensors 3-4 matrices sensitive to visible infrared. In other words, the analysis unit 5 inputs at a predetermined frequency an image taken by the sensor 3 and an image taken by the sensor 4 (step 20 in Figure 4). These two images taken synchronously form a pair of images of the environment 10; detection of the objects in motion in the environment 10 from a sequence of successive images acquired from one of the two cameras 3-4 (the camera 3 in the example of FIG. 4) (step 21 in FIG. 4 ): in case of non-detection of moving objects (test 22 in FIG. 4), then resuming the process in the initial step (arrow returning to step 20 for a new iteration of the process), otherwise the process continues ; 2D tracking of the detected objects in motion (step 23 in FIG. 4); skeletonization and discretization of the objects tracked (step 24 in FIG. 4); mapping of the discrete points of the objects followed with those included in the corresponding image taken by the second sensor (the camera 4 in the example of FIG. 4): search in the corresponding image rendered by the sensor 4 of the points counterparts of the discrete points of the skeletons of the moving objects detected in the image taken by the sensor 3 (step 25 in FIG. 4); calculating the 3D positions, by triangulation, of the discrete points of the skeletons of the tracked objects, the two cameras 3-4 being previously calibrated spatially with respect to the ground of the monitored environment (step 26 in FIG. 4); analysis of the vertical distribution of the discrete points of each skeleton: plotting of the average height over time (that is to say, issuing from several images in time) of the points of each skeleton (step 27 in FIG. ). The comparison with predefined values of the speed of variation of the average heights of these discrete points makes it possible to detect a fall, the latter corresponding to a rapid decrease in the average height of an object followed by a delay on the ground (this is ie the fall is confirmed if the object remains a certain time on the ground) (test 28 in Figure 4). At the same time, we proceed to the next iteration (arrow pointing to step 20) by receiving synchronously a new pair of images rendered by the two cameras 3-4; in the event of detection of a fall (test 28 in FIG. 4), signaling of this fall via the communication module 6 (step 29 in FIG. 4). The detection of moving objects in the environment 10 from a sequence of successive images taken by the sensor 3 is made on the basis of the variations in the time of the contours (Sobel algorithm) of an image at the next in this sequence, a sequence of images being formed by a predetermined number of the latest images acquired from the infrared visible sensitive matrix sensor 3.

The detection of moving objects from a sequence of images acquired from the sensor 3 proceeds by two preprocessing steps, namely, - the suppression of the shadows: in day and night scenes, the lighting (in inside or outside) often projects shadows fairly pronounced, on the floor, on the wall or on objects around. Indeed, if a person passes in front of the light source, a classic motion detection tends to detect the shadows as moving objects. However, thanks to the device 7 infrared lighting, which has the effect of reducing shadows because it is oriented towards a diffuser (the ceiling for example), the borders of these shadows become quite soft and are no longer or not very visible during edge detection. Therefore, the approach taken is to do motion detection based on the variations of the image contours over time because none of the shadows generates an outline even with relatively low detection thresholds (see figure 5); calculation of the background (subtraction of the background): the background is defined in a sequence of images as the set of still pixels in the image sequence for a predefined duration, called history. An edge detection (Sobel algorithm) is then applied to the image and then the background of the contours is created as follows: if a pixel is detected as an outline on more than half of the history of contours, then this pixel is part of the contours of the background. Finally, a difference between the current contours and the background of the contours makes it possible to find the outlines in motion (image b of FIG. 9). The morpho-mathematical operators of opening and closing are applied on the image obtained to connect the close objects and to suppress the detection noise (image c of FIG. 9). Advantageously, the infrared lighting device 7 of the monitored environment has the effect of attenuating the shadows and thus of improving the accuracy of detection of an object moving in the environment 10. The 2D tracking step detected objects moving in the environment 10 is intended, in particular, to analyze over time the behavior of any distinguished person and to detect any new person who appears in this environment 10.

For this purpose, the 2D tracking of the objects detected while moving in the environment 10 is obtained in the following manner: smoothing of the detected contours; - identification of related components in these smoothed outlines; - clustering of related components: in other words the distinction of objects (these objects being assumed to be "people"); and - tracking the movements of the distinguished objects based on the modeling of the fixed background (each distinguished person is represented by a binary mask). Indeed, in a first step, the connected components (CC) are identified in the mask representing each pixel with movement. These CCs are then filtered according to their size to eliminate the noise: if a CC is composed of less than a threshold expressed in pixels, then it is part of the noise. Otherwise, the following algorithm is used to aggregate CCs to objects (ie to people): - CCs are sorted by size (number of pixels); - for each CC, a score is calculated for each existing object, based on: o the proximity of the CC with the object to the previous image; o the difference in size between the CC and the object in the previous image - if the CC score is below a certain threshold, then it is aggregated to the object. Otherwise, a new object is created from the CC: this means that a new person has entered the field, which must also be followed.

Thus, each moving object in the field of the two cameras 3-4 is followed, and represented by a binary mask. The next phase is the skeletonization / discretization of the objects tracked: in mathematical morphology, the skeleton of a binary object is defined as the set of centers of the maximum balls contained in this binary object. The skeleton obtained is then discretized at a number of points, as shown in FIG. 7. In a next step, the 3D attitude of the persons detected in the environment 10 is reconstructed on the basis of the stereovision principle and the epipolar geometry. It should be noted that stereovision consists in using two synchronous images of the scene. With reference to FIG. 8, a point P of the scene is seen in two dimensions in each of the two images rendered by the cameras 3 and 4.

Knowing the calibration parameters of the system, the 3D position of the point P can be calculated by triangulation from this pair of positions p1 and p2 of the point P seen from the two cameras 3-4. The principle of the epipolar constraint can be deduced from this geometric diagram (see Figure 8), noting that the points C1, C2 (optical centers of the cameras 3-4), pl, p2 (projections of the point P on the images, Image_1 and Image_2, both cameras 3-4) and the point P itself are coplanar. The intersection of this plane with the plane of each of the images is a line, called the epipolar right. Thus, the homologue of a point p1 on the image Image_l is necessarily, in the image Image_2, on the epipolar line formed by the intersection of the plane (C1, C2, pl) and the image Image_2. This constraint facilitates the mapping of points from one image to another by pattern recognition in terms of robustness and computation time. In one implementation, the mapping and triangulation is obtained by searching at each instant, in the image rendered by the camera 4, the counterparts of all the discrete points of the skeleton of the or each moving object detected in the image of the camera 3 (arrow connecting step 20 to step 25 in FIG. 4). This step of matching the points by shape recognition is performed by normalized cross-correlation along the epipolar line of each of the points of each skeleton identified in the image rendered by the camera 3. We deduce the 3D positions of the points matched by triangulation (see Figure 8) based on the spatial calibration performed at the installation of the fall detection system 1. The 3D coordinates of these skeleton points are expressed in the spatial landmark of the area under surveillance. Figures 9 and 10 illustrate the distribution of the 3D points of a skeleton, respectively, of a standing person and a person lying on the ground.

Now, from the 3D points obtained from the skeleton in the ground reference of the monitored zone, the attitude of the moving object is analyzed. In particular, its height is plotted over time, as shown in FIG. 11. To limit the noise, the average values over time of the height of an object are used (the smooth curve of FIG. 11). From this curve, it is considered that a fall corresponds to a sudden variation of the average height of an object downwards followed by a delay on the ground. The decay rate of the average height of a person in normal activity generally remains below a certain threshold whereas during a fall it exceeds this limit. The detection of falls of a person from a curve representative of the variation of its average height in a sequence of successive images (and therefore over time) is based on detecting in this curve a signature of fall . For example, the scenario "relatively high average height (Standing) - decreasing average height at a fast speed - average height near zero (elongated)" represents a fall signature, however the scenario "average height relatively large (Standing), decay of the average height at an average speed - average height slightly greater than zero (elongated) "represents a lengthening signature and not that of a fall. By way of example, in FIG. 11, the last fall corresponds to a seated person who, when getting up, falls to the ground (therefore, the fall from a sitting position is taken into account by the present fall detection system. ). Advantageously, the thresholds mentioned above, being variable from one person to another and from one configuration to another configuration of the digital camera 2, are adjustable. The communication module 6 is configured to inform an external module (not shown in FIG. 1) of a fall of a person occurring in the monitored environment. This plug-in can be a remote server of a medical platform, a user terminal, a remote assistance center, or more generally any other means of communication. The communication module 6 is provided with a plurality of communication interfaces capable of transmitting information via communication networks (an SMS, an MMS, a direct transmission of images from the digital camera system, an email, a fax, a call for example) to a predefined recipient. The communication module 6 is further configured to respond to a request received via one of its communication interfaces (request to transmit live images of the environment 10, query relating to statistics relating to the falls and / or those present. in the environment 10 for example). In one embodiment, the information transmitted by the communication module is coded.

Advantageously, the method described above also makes it possible to locate the place of the fall of a person in the environment 10 (a fall on the track in a subway station, fall on the bed, fall on the stairs by example) to trigger an alert / alarm in case of a risk situation. This geolocation is assured of course thanks to the preliminary calibration step during which we can distinguish and name different areas in the spatial reference linked to the environment 10. The method just described has a number of advantages. In fact, infrared lighting makes it possible to provide diffused lighting in the environment to be monitored and, consequently, to avoid the projection of shadows in this environment. This has the effect of having a 3D representation more faithful to the real mobile object and, therefore, a more precise tracking of this object in the monitored environment; the bi-camera digital imaging device 2 (ie only having two cameras with almost the same field of view, as opposed to conventional fall detection systems which include more than two spatially distributed cameras ) has the advantage of being compact and allowing the adoption of the 3D reconstruction method presented above which can not be used with a system composed of several cameras having very different fields of view; the subtraction of the background following the detection of the smoothed outlines of objects has the effect of detecting the movement by tolerating movements of objects. Advantageously, according to the above-described method and system for detecting falls, only two cameras are required to monitor and detect the fall of one of a plurality of persons present in an environment included in the field of view of these two cameras. . The fall detection system described above finds particular application in a retirement home, in a hospital (waiting room, corridors, entrance / exit for example), in a metro station, or more generally any place intended for dependents and / or people exposed to falls. It should be noted that the terms "module" or "unit" here cover any physical box incorporating a processor programmed to perform one or more predetermined functions, and / or any software application (program or subroutine, plugin) implemented on a processor independently or in combination with other software applications.

Claims (10)

CLAIMS1 A system (1) for detecting falls of persons comprising - a digital camera (2) provided with a first sensor (3) sensitive matrix visible infrared and a second sensor (4) visible matrix sensitive to the infrared (4), the two sensors (3-4) looking substantially differently the same scene (10); an analysis unit (5) programmed to acquire from said digital imaging device (2) an image taken by said first infrared visible sensitive matrix sensor (3) and an image taken by said second sensor (4) sensitive matrix visible infrared, these two images being taken synchronously and forming a pair of images of said scene (10); o detecting the fall of at least one person in said scene (10) from a sequence in time of image pairs of said scene (10). 2. A system according to the preceding claim, characterized in that it further comprises an infrared lighting device (7) arranged to provide diffuse illumination in the scene (10). 3. A system according to claim 1 or 2 characterized in that it further comprises a communication module (6) allowing said system to interact with other communication systems. 4. A system according to any one of the preceding claims, characterized in that the first and second sensors (3-4) matrices sensitive to the visible infrared are spatially calibrated relative to the ground of said scene (10). 5. A system according to any one of the preceding claims, characterized in that the analysis unit (5) is programmed to detect moving objects in said scene (10) from the images taken by said first sensor (3) matrix sensitive to visible infrared; tracking in two dimensions the objects detected in the images taken by said first sensitive matrix sensor (3) from the visible to the infrared; skeletonizing, in at least one image taken by said first sensitive matrix sensor (3) from visible to infrared, objects tracked in two dimensions; - discretize at least one skeleton; searching in the image taken by said second sensitive infrared and corresponding sensitive matrix sensor (4), in a pair of images, at said at least one image taken by said first sensitive visible matrix sensor (3) infrared points homologous to the discrete points of the discretized skeleton; calculating the 3D positions of the discrete points of said skeleton; - Compare with predefined values the rate of variation over time of the average heights of the discrete points. 6. A system according to the preceding claim, characterized in that the unit (5) of analysis, for detecting moving objects, is arranged to use the variation in time of the contours of one image to the next in a sequence of successive images taken by said first sensor (3) matrix susceptible to visible infrared. 7. A system according to claim 5, characterized in that the analysis unit (5) for tracking detected objects in two dimensions is arranged to smooth out the detected contours; - identify related components in smoothed outlines; - to distinguish objects on the basis of the identified connected components; - to represent each object distinguished by a binary mask. 8. A system according to claim 5, characterized in that the analysis unit (5) for calculating the 3D positions uses the principle of epipolar averaging for a pair of images taken from two substantially different angles of view. 9. A system according to claim 3, characterized in that said communication module (6) comprises a plurality of communications interfaces. A method of detecting falls of persons comprising the steps of: synchronously taking a pair of digital images of a scene (10), an image at a first angle of view and an image in a second angle of view, the first and the second angle of view being substantially different; detecting the fall of at least one person in said scene (10) from a time sequence of digital image pairs of said scene (10). 15 1 1. A method according to the preceding claim, characterized in that it further comprises a step of diffuse infrared illumination of said scene (10). 12. A method according to claim 10 or 11, characterized in that it further comprises a signaling step of detecting the fall of a person. A method according to claim 10, characterized in that the step of detecting the fall of at least one person comprises - detecting moving objects in said scene (10) from the images taken at the first angle from view; The two-dimensional tracking of the objects detected in the images taken under the first angle of view; skeletonization, in at least one image taken at the first angle of view, objects tracked in two dimensions; the discretization of at least one skeleton; The search in the image taken under said second angle of view and corresponding, in a pair of images, to said at least one image taken at the first angle of view of the points homologous to the discrete points of the discretized skeleton; calculating the 3D positions of the discrete points of said skeleton; the comparison, at predefined values, of the speed of variation over time of the mean heights of the discrete points. 14. A method according to claim 13, characterized in that the step of detecting moving objects uses the variation, in time, of the outlines of one image to the next in a sequence of successive images taken under said first angle of view. 15. A method according to claim 13, characterized in that the 2D tracking step comprises a step of smoothing the detected contours; a step of identifying the connected components in the smoothed outlines; a step of distinguishing objects based on the identified connected components. a step of representing each object distinguished by a binary mask. 16. A method according to claim 13, characterized in that the calculation of 3D positions uses the principle of epipolar geometry for a pair of images taken from two substantially different angles of view. 17. Computer program product implemented on a memory medium, capable of being implemented within a computer processing unit and comprising instructions for the implementation of a method according to one of claims 10 to 16.