Method and system for tracking moving objects in a sequence of occultation-sensitive images

Abstract

A computer-implemented method for tracking a moving object in a sequence of images; the method takes as input a sequence of images and provides as output the position of a moving object within the sequence of images.The method includes the following steps: - detect a moving object in a first image of the image sequence; - predict the position of said moving object in at least a second image of the image sequence, the position being a bounding box; - calculate a community coefficient between the predicted bounding box corresponding to said moving object in the second image and the bounding box corresponding to an object detected in the second image; - validate the predicted position of the moving object if the value of the community coefficient is greater than a threshold value; in which the community coefficient CC is the ratio of the area of ​​the intersection between the predicted bounding box and the bounding box corresponding to a detected object to the area of ​​their union, the area of ​​each of the two bounding boxes being an area weighted according to a surface distribution law previously established for each of the images of the image sequence.

Description

Title of the invention: Method and system for tracking moving objects in a sequence of occultation-sensitive images. Technical field

[0001] The invention relates to a method and a system for tracking moving objects in a sequence of images sensitive to occultation. Technical background

[0002] The detection and tracking of moving objects, such as people, animals, or vehicles, in a sequence of images or a video are essential operations in computer vision. The applications are numerous. For example, these operations form the basis of any analysis of interactions between objects and their environment for the operation of so-called autonomous systems such as autonomous vehicles or robots, for the rational management of crowds or road traffic, and for statistical studies of modeling and predicting collective behaviors of individuals. They are also used in railway and airport gates for the registration, identification, and / or biometric authentication of travelers or vehicles in free-flowing traffic.As part of civil security prevention efforts, police investigations, or the fight against crime, they also provide essential information for a better understanding and interpretation of everyday life scenes, offenses, or crimes.

[0003] The tracking of moving objects in a sequence of images or a video generally relies on statistical learning methods such as artificial neural networks. These methods are configured to determine the trajectory of one or more objects through a sequence of images and consist, in globo, of detecting one or more types of objects between successive images of an image sequence, delimiting them and / or predicting their position using bounding boxes, and then evaluating the overlap rate of the bounding boxes between successive images.

[0004] It is common to evaluate the degree of overlap between two bounding boxes using the Jaccard index, also called the "Jaccard coefficient" or "intersection over union" (LOU). The Jaccard index is insensitive to scale effects. Its value ranges from 0 to 1. A value of 0 means that the two sets A and B are perfectly disjoint, and a value of 1 means that they perfectly overlap. It is defined as the ratio of the cardinality of the intersection between two sets A and B and the cardinality of the union between these two sets. J(4 B) = --

[0005] For two bounding boxes between two successive images, the Jaccard index corresponds to the ratio of the area of ​​the intersection between the two bounding boxes to the area of ​​their union. In the context of tracking moving objects in a sequence of images or in a video, two bounding boxes between two successive images are generally considered to refer to the same object when the Jaccard index is greater than a predefined threshold value.

[0006] In practice, as long as the Jaccard index value of the bounding boxes between successive images of a sequence or video remains above this threshold value, the bounding boxes are assumed to represent the same object between each image. The object is then tracked in the image sequence or video. Conversely, as soon as the Jaccard index value falls below this threshold value, the two bounding boxes are assumed to represent two different moving objects. Object tracking then ceases.

[0007] US 2020134837 Al [INTEL CORP [US]] 30.04.2020 describes, for example, a method for tracking moving objects in a video in which a Jaccard index is calculated between blobs and bounding boxes representing the objects detected by a convolutional neural network over successive frames. If the value of the Jaccard index is less than a threshold value, the convolutional neural network performs a new object detection. This method makes it possible, in particular, to detect new moving objects entering a scene.

[0008] WO 2020 / 091749 Al [ARCUS HOLDING AS [DK]] 07.05.2020 describes a method for detecting moving objects in a video by simultaneously combining a deep learning approach in the form of a convolutional neural network and a pixel value change detection approach. Objects detected by these two approaches are bounded by bounding boxes. A Jaccard index is calculated between the bounding boxes of each of the two approaches and then compared to a threshold value.

[0009] The effectiveness of approaches based on comparing the Jaccard index value to a threshold value relies on the assumption that a moving object is always fully, or almost fully, visible from one image to the next. However, in a real-world environment, it is common for a moving object to be partially obscured by other fixed or moving objects in its environment during its movements. For example, an individual initially fully visible in a first image may be partially obscured in the next image by other individuals behind whom it is moves or by a vehicle moving in front of it. Similarly, a vehicle initially visible in its entirety in a first image may be partially obscured in the next image when it is overtaken by another vehicle or obscured by the movement of a crowd of pedestrians.

[0010] A first negative consequence is the untimely and erroneous interruption of the tracking of a moving object. During an occultation, the value of the Jaccard index can become artificially lower than the threshold value. Since the bounding boxes of two successive images are then no longer considered representative of the same moving object, the tracking is interrupted.

[0011] A second negative consequence is the confusion between two moving objects, resulting in the tracking of a different moving object than the one initially being tracked. When a first moving object is occulted by a second moving object of similar nature and size, for example, two vehicles, the bounding box corresponding to the second moving object may be mistakenly considered to be the bounding box of the first moving object. The value of the Jaccard index may artificially become greater than the threshold value, and the tracking may continue erroneously based on the second moving object.

[0012] To overcome these drawbacks, various strategies have been developed in which the Jaccard index is combined with several other indices, parameters, scores or information obtained using other image processing methods.

[0013] WO 2020 / 175818 Al [NAVER CORP [KR]] 03.09.2020 describes a method for tracking moving objects in an image sequence based on a classification-type network configured to calculate a posture similarity index between objects in successive images. This similarity index improves the detection and tracking of occulted objects. It includes a displacement factor calculated based on a Jaccard index and an appearance factor based on a measure of correspondence between global patterns.

[0014] WO 2021 / 242445 Al [MICROSOFT TECHNOLOGY LICENSING LLC [US]] 02.12.2021 describes a method for tracking moving objects in a video capable of handling partial or total object occlusion. The method implements a detection module and a tracking module that simultaneously process the images of a video stream. An association module allows the objects detected by the detection module to be associated with the objects tracked by the tracking module via a spatial affinity score based on the Jaccard index. The tracking module is further configured to anticipate the trajectory of a moving object based on information provided by the detection module.

[0015] WO 2021 / 237749 Al [SIEMENS AG [DE]] 02.12.2021 describes a method for tracking moving objects, such as vehicles, in a sequence of images of a Video. Based on the premise that different objects have different sizes, it implements, in addition to the Jaccard index, a size change index between the bounding boxes of two consecutive images. If the size change index is less than a given threshold value and the Jaccard index is greater than a given value alone, the bounding boxes of the two successive images are considered to represent the same object.

[0016] WO 2022 / 256150 Al [MICROSOFT TECHNOLOGY LICENSING LLC [US]] 08.12.2022 describes a method for tracking a plurality of moving objects, such as individuals, in a video image sequence. The method is based on the assumption that the same object between two consecutive images retains its appearance and that different objects have different sizes. The method implements a neural network trained to predict and estimate the position and size of the occlusion zones between objects during their movement by anticipating their trajectory. This predictor / estimator of occlusion zones can be combined with a Jaccard index. Summary of the invention Technical problem

[0017] A major drawback of current methods for tracking moving objects capable of handling their temporary and / or partial occultation by other objects in a scene is the complexity of their implementation, in particular the need to implement other additional image processing methods and combine them with a Jaccard index calculation, which consequently requires additional computing resources.

[0018] Another drawback of these methods is that, by attempting to overcome only the inherent limitations of the Jaccard index for detecting partially or totally obscured moving objects, they remain dependent on it. In other words, they remain based on an index which, as defined and used therein, is unsuitable for addressing the occultation phenomena of moving objects that can occur in a sequence of images of a real scene.

[0019] There therefore remains a need for a simple and robust method for tracking moving objects capable of handling their temporary and / or partial occultation. Technical solution

[0020] According to a first aspect of the invention, a computer-implemented method is provided for tracking a moving object in a sequence of images. The method takes as input data a sequence of images and provides as output data the position of a moving object within the sequence of images. The method comprises the following steps: - detect a moving object in the first image of the image sequence; - predict the position of said moving object in at least one second image of the image sequence, the position being a bounding box; - calculate a community coefficient between the predicted bounding box corresponding to said moving object in the second image and the bounding box corresponding to an object detected in the second image; - validate the predicted position of the moving object if the community coefficient value is greater than a threshold value; in which the community coefficient is the ratio of the area of ​​the intersection between the predicted bounding box and bounding box corresponding to a detected object to the area of ​​their union, the area of ​​each of the two bounding boxes being an area weighted according to a surface distribution law previously established for each of the images in the image sequence.

[0021] Advantageous embodiments are described below.

[0022] According to a second aspect of the invention, a data processing device is provided comprising means for implementing the method in any of the embodiments of the first aspect of the invention.

[0023] According to a third aspect of the invention, a computer program is provided comprising instructions which, when the program is executed by a computer, lead the computer to implement the method according to any one of the embodiments of the first aspect of the invention.

[0024] According to a fourth aspect of the invention, a storage medium is provided comprising instructions which, when executed by a computer, cause the computer to implement the method according to any one of the embodiments of the first aspect of the invention.

[0025] According to a fifth aspect of the invention, a system is provided Brief description of the drawings

[0026] [Fig. 1] is a schematic representation of a first image and a second image of a scene in which a moving object is visible in its entirety and then partially obscured during its movement.

[0027] [Fig.2] is a flowchart of a method according to the first aspect of the invention.

[0028] [Fig.3] is a schematic representation of a surface distribution law for the second image of the scene in [Fig.1]

[0029] [Fig.4] is a schematic representation of a data processing device according to the second aspect of the invention. Detailed description of the implementation methods

[0030] Figure 1 shows an example of a two-image sequence II, 12 of a scene in which two individuals 1001, 1002 are moving. The two individuals 1001,1002 are respectively represented by bounding boxes B1-1001 and B1-1002 on the first image II and by bounding boxes B2-1001 and B2-1002 on the second image 12. These bounding boxes can be obtained by any type of image analysis method capable of detecting objects, including convolutional neural networks.

[0031] In the first image II, the two mobile individuals 1001, 1002 are far apart and their bounding boxes Bl-1001, Bl-1002 do not intersect. The overlap rate, calculated using a Jaccard index, between the two bounding boxes Bl-1001, Bl-1002 is therefore zero.

[0032] In the second image 12, the two individuals 1001 and 1002 cross paths, and the bounding boxes B2-1001 and B2-1002 overlap. The overlap rate, calculated using a Jaccard index, between the two bounding boxes B2-1001 and B2-1002 is higher than in the case of the first image II. If this rate exceeds a threshold value, there may be a risk of confusion between the two individuals by a conventional tracking system. The system may consider the two individuals 1001 and 1002 to represent the same individual, for example, individual 1002, and stop tracking individual 1001.

[0033] In addition to detecting moving objects and delimiting them using bounding boxes, a moving object tracking system can be configured to predict the position of a moving object as it moves through one or more images in an image sequence. This prediction can be established by extrapolation from a displacement field whose measurement is based on an analysis of the different positions of the object and / or one or more distinctive characteristics of this moving object from one or more previous images.

[0034] By way of example, with reference to [Fig. 1], the tracking system can predict the position of the first individual 1001 from a displacement field measured by analyzing the different positions of its head and / or upper body, such as its torso, possibly combined with an analysis of its build and / or posture. In the first image II, since the first individual 1001 is fully visible, the predicted box PI-1001, representing the predicted position, corresponds more or less to the encompassing box Bl-1001. In contrast, in the second image 12, the predicted position is represented by a box P2-1001 centered only on a visible part of individual 1001, in this case, the upper part of its torso, including its head. The predicted box P2-1001 overlaps the bounding box B2-1001 of the first character 1001 as well as part of the bounding box B2-1002 of the second character 1002.

[0035] For tracking systems that are wholly or partly based on the Jaccard index, The predicted position is validated when the value of the index between the predicted box and the bounding box of the moving object is greater than a threshold value, for example, 0.7. In image II, this condition is met: the Jaccard index between the predicted box Pl-1001 and the bounding box Bl-1001 is greater than 70%. In image 12, the value of the Jaccard index between the predicted box P2-1001 and the bounding box B2-1002 representing individual 1002 is less than 70%. The condition is not met; individual 1001 is not confused with individual 1002. However, in image 12, the value of the Jaccard index between the predicted box P2-1001 and the encompassing box B2-1001 of individual 1001 is also less than 70%, the condition is not met for the first individual 1001 either, and the predicted position P2-1001 is not validated for individual 1001. Yet, this predicted position P2-1001 does indeed correspond to the first individual 1001.Tracking systems that are wholly or partly based on the Jaccard index therefore fail to track moving objects that may be obscured.

[0036] According to a first aspect of the invention, with reference to [Fig. 1] to 3, a computer-implemented method 2000 is provided for tracking a moving object 1001 in a sequence of images II, 12. Method 2000 takes as input data a sequence of images II, 12 and provides as output data the position P1-1001, P2-1001 of a moving object 1001 within the sequence of images II, 12. Method 2000 comprises the following steps: - detect 2001 a moving object 1001 in a first image II of the image sequence II, 12; - predict 2002 the position P2-1001 of said moving object 1001 in at least a second image 12 of the image sequence II, 12, the position P2-1001 being a bounding box; - calculate 2003 a community coefficient CC between the predicted bounding box P2-1001 corresponding to said moving object 1001 in the second image 12 and the bounding box Bl-1001, B2-1002 corresponding to a detected object 1001, 1002 in the second image 12; - validate 2004 the predicted position P2-1001 of the moving object 1001 if the value of the community coefficient CC is greater than a threshold value VS; in which the community coefficient CC is the ratio of the area of ​​the intersection between the predicted bounding box P2-1001 and bounding box Bl-1001, B2-1002 corresponding to a detected object 1001, 1002 to the area of ​​their union, the area of ​​each of the two bounding boxes being an area weighted according to a surface distribution law G previously established for each of the images II, 12 of the image sequence II, 12.

[0037] The community coefficient is defined as the ratio of the area of ​​the intersection between the predicted bounding box P2-1001 and the bounding box Bl-1001, B2-1002 corresponding to a detected object 1001, 1002 on the area of ​​their union, Faire of each of the two bounding boxes being an area weighted according to a surface distribution law G previously established for each of the images II, 12 of the image sequence II, 12.

[0038] A fundamental difference between the community coefficient according to the invention and the Jaccard index (loU, "Intersection over Union") is the weighting of the areas of the bounding boxes according to a surface distribution law. The function of this weighting is to ensure a differentiated contribution of the regions of the image contained in the bounding boxes according to their relevance to the characteristics of the moving object they represent.In other words, image regions representing relevant or distinguishing features of the moving object, such as, for example, the head or upper torso in the case of an individual, are given priority in their contribution to the area of ​​the bounding box in which they are contained. A notable advantage of the community coefficient according to the invention is therefore improved discrimination of moving objects, particularly with regard to the features to be taken into account to differentiate them during occlusion phenomena.

[0039] Figure 3 shows a schematic and purely illustrative example of a surface distribution law in the form of a point gradient pattern G for the second image 12 of Figure 1. The three bounding boxes P2-1001, B1-1001, and B2-1002 are also shown in this figure. The point gradient pattern G consists of three zones G1, G2, and G3 whose point density decreases from the top of the image to the bottom.These three zones Gl, G2, G3 represent a surface distribution law corresponding to decreasing weightings of pixels from the top of the image to the bottom of the image when calculating the areas of the three bounding boxes P2-1001, Bl-1001 and B2-1002: during this calculation, the pixels of the bounding boxes located in the first zone Gl are weighted by a weighting coefficient higher than those of the second zone G2 which are themselves weighted by a weighting coefficient higher than those of the third zone G3.

[0040] Thanks to this surface distribution law, pixels have different weights when calculating the area of ​​bounding boxes depending on their location in the image. By analogy, in the example illustrated in [Fig. 3], the centroids of the predicted bounding box P2-1001 and bounding box Bl-1001 are respectively shifted towards the area of ​​the distribution in which the weighting of the pixels is highest, in this case the first area Gl; the pixels located in the first area Gl have more importance than those located in the second area G2.

[0041] During the calculation of the areas of the predicted bounding box P2-1001 and the bounding box Bl-1001 corresponding to individual 1001, the pixels of the first area Gl are weighted by a coefficient higher than those of the second zone G2. The value of the community coefficient according to the invention between the predicted bounding box P2-1001 and the bounding box B1-1001, being more representative of this weighting, becomes higher than that of a Jaccard index, and, unlike the latter, when individual 1001 is occulted by individual 1002, remains higher than the threshold value VS. The predicted position for individual 1001 is validated; the tracking of individual 1001 continues correctly despite the occultation.

[0042] In the example in [Fig. 3], the surface distribution law is intentionally simplified for purely illustrative purposes. Thus, instead of being discontinuous, particularly organized into regions of different density, it is preferably continuous across the entire image. Similarly, the variations in weighting along the image may differ from that shown in [Fig. 3], which is decreasing from the top to the bottom of the image. For example, it may be decreasing or increasing from one side of the image to another, or from the edges of the image to its center. In general, whether the surface distribution law is continuous or discontinuous, as well as its variations along the image, depend on the type of moving objects that may be tracked, how they move, and the environment in which they move.For example, they may differ depending on whether the moving object is a person or a vehicle. The surface distribution law may be different for each image in the image sequence. When the nature of the scene in which the moving objects are tracked does not change substantially during tracking, the surface distribution law is generally identical for each image in the image sequence.

[0043] The surface distribution law can be of any suitable type. It can be in the form of a table of values ​​in which each value is a weighting coefficient associated with a region comprising one or more pixels of the images in the image sequence. It can also be in the form of a mathematical function whose value varies according to the location of the pixels of the images in the image sequence.

[0044] According to certain preferred embodiments, the surface distribution law G is an exponential distribution law. An exponential distribution is advantageous in that it allows for differentiated weighting, granting a significantly high contribution to certain most relevant regions of an image while considerably reducing the contribution of less relevant regions in the calculation of the community coefficient according to the invention. Such a distribution is particularly well-suited for tracking individuals in a scene.

[0045] According to an example, for a decreasing exponential distribution law from the top to the bottom of the image, the area A of a bounding box can be calculated using the following formula, where 1 is the width, expressed in pixels, of the bounding box, y

[0046]

[0047] the coordinate of the top left corner of the box, h the height, expressed in pixels, of the bounding box and T a fixed parameter. A = r eiT = er?7^ An exponential distribution law or similar distribution may make the community coefficient according to the invention sensitive to scaling effects such as enlarging or reducing the images in the image sequence. Since such operations may be performed on the images, this sensitivity can be a source of artifacts that may manifest as aberrant values ​​in the sensitivity coefficient.

[0048] Also, according to certain advantageous embodiments, the exponential distribution law may further include, as a parameter, a normalization factor with respect to at least one dimension of the images in the image sequence. This normalization factor serves to eliminate sensitivity to scale effects. The normalization may be performed with respect to the image width, the image height, or the dimensions.

[0049] From the formula of the previous example, the fixed parameter T can include a normalization factor expressed as follows, with U being the height of the image expressed in pixels: 100501

[0051] In the preceding formula, the normalization factor corresponds to the coefficient 1 / U.

[0052] The parameter T here functions as an empirically defined optimization parameter. It can be a fixed or variable number, preferably fixed. It has been found that values ​​between 0 (not included) and 30 (inclusive) for the parameter T constitute an optimal range of values ​​for tracking individuals in a sequence of images in which they are likely to be obscured.

[0053] In general, the distribution law may advantageously include an optimization factor, of which the preceding parameter T is an example. This parameter may be a fixed or variable value, preferably a fixed value. Its value may be adjusted using a purely empirical approach or with the aid of any suitable optimization method.

[0054] According to some preferred embodiments, the surface distribution law G includes, as a parameter, an optimization factor, the optimization factor being previously defined by minimizing the difference between the community coefficient and a Jaccard index for unocculted moving objects. An optimization factor obtained by minimizing the difference between the coefficient of community according to the invention and a Jaccard index for non-occulted moving objects is particularly advantageous in that it allows, during its implementation, a community coefficient that preserves the continuity of tracking a moving object when it is occulted and maintains a behavior close to that of a Jaccard index when the object is no longer occulted.

[0055] The method according to the invention is particularly advantageous for tracking moving objects such as individuals or vehicles in a sequence of images. Thus, according to preferred embodiments, the moving object is an individual in a substantially vertical position, and the surface distribution law G is a vertically decreasing distribution from the head to the feet of upright individuals. According to other embodiments, the moving object is a vehicle, and the surface distribution law G is a horizontally decreasing distribution in the direction of vehicle traffic.

[0056] The method according to the first aspect of the invention is implemented by computer. With reference to [Fig. 4], in a second aspect of the invention, a data processing device 4000 is provided, comprising means for implementing a method 2000 according to any one of the embodiments of the first aspect of the invention.

[0057] An example of a device may be one designed to automatically execute sequences of arithmetic or logical operations to perform tasks or actions. This device, also called a computer, may include one or more central processing units (CPUs) and / or one or more graphics processing units (GPUs) 4001, as well as at least one control device adapted to perform these operations. It may also include other electronic components such as input / output interfaces 4002, non-volatile or volatile storage devices 4003, and communication buses for transferring data between internal components of the device or with external components. One of the input / output devices 4002 may be a user interface for human-machine interaction, for example, a graphical user interface for displaying human-understandable information.

[0058] According to a third aspect of the invention, a computer program 14003 is provided comprising instructions which, when the program is executed by a computer, lead the latter to implement a method 2000 according to any one of the embodiments of the first aspect of the invention.

[0059] Any type of programming language, compiled or interpreted, can be used to implement the steps of the method of the invention. The computer program may be part of a software solution, that is to say, a collection of executable instructions, code, scripts or other elements, and / or databases.

[0060] According to a fourth aspect of the invention, a computer-readable recording medium 4003 is provided comprising instructions which, when executed by a computer, cause the computer to implement a method 2000 according to any one of the embodiments of the first aspect of the invention.

[0061] The computer-readable recording medium 4003 is preferably non-volatile memory, for example a hard disk drive or a solid-state drive. It may be removable storage media or non-removable storage media that is part of a computer.

[0062] The computer-readable recording medium 4003 can also be volatile memory within a removable medium. This can facilitate the deployment of the invention at numerous production sites.

[0063] The computer-readable recording medium 4003 may be part of a computer used as a server from which executable instructions may be downloaded and, when executed by a computer, cause the computer to execute a method according to one of the embodiments described in this document.

[0064] The computer program 14003 and the storage medium 4003 on which it is recorded can be implemented in a distributed computing environment, for example, cloud computing. The instructions can be executed on a server to which one or more client computers can connect and provide encoded data as input data for a method according to any one of the embodiments of the first aspect of the invention. Once the data has been processed, the result can be downloaded and decoded on the client computer or sent directly, for example, in the form of instructions.

[0065] According to a fifth embodiment, a system for tracking a moving object in a sequence of images is provided. The system comprises - an image acquisition device; - a data processing device according to the second aspect of the invention, said data processing device being configured to receive and process a sequence of images acquired by the image acquisition device. References Literature patent

[0066] US 2020134837 Al [INTEL CORP [US]] 04.30.2020.

[0067] WO 2020 / 091749 Al [ARCUS HOLDING AS [DK]] 07.05.2020.

[0068] WO 2020 / 175818 Al [NAVER CORP [KR]] 03.09.2020.

[0069] WO 2021 / 242445 Al [MICROSOFT TECHNOLOGY LICENSING LLC [US]] 02.12.2021.

[0070] WO 2021 / 237749 Al [SIEMENS AG [DE]] 02.12.2021.

[0071] WO 2022 / 256150 Al [MICROSOFT TECHNOLOGY LICENSING LLC [US]] 08.12.2022.

Claims

Demands

1. A computer-implemented method (2000) for tracking a moving object (1001) in a sequence of images (II, 12), the method (2000) takes as input data a sequence of images (II, 12) and provides as output data the position (P1-1001, P2-1001) of a moving object (1001) within the sequence of images (II, 12), the method (2000) comprises the following steps: - detecting (2001) a moving object (1001) in a first image (II) of the sequence of images (II, 12); - predicting (2002) the position (P2-1001) of said moving object (1001) in at least a second image (12) of the sequence of images (II, 12), the position (P2-1001) being a bounding box; - calculate (2003) a community coefficient CC between the predicted bounding box (P2-1001) corresponding to said moving object (1001) in the second image (12) and the bounding box (B 1-1001, B2-1002) corresponding to a detected object (1001, 1002) in the second image (12);- validate (2004) the predicted position (P2-1001) of the moving object (1001) if the community coefficient CC is greater than a threshold value VS; where the community coefficient CC is the ratio of the area of ​​the intersection between the predicted bounding box (P2-1001) and the bounding box (B1-1001, B2-1002) corresponding to a detected object (1001, 1002) to the area of ​​their union, the area of ​​each of the two bounding boxes being an area weighted according to a surface distribution law G previously established for each of the images (II, 12) of the image sequence (II, 12).;

2. Method (2000) according to claim 1, such that the surface distribution law G is an exponential distribution law.

3. Method (2000) according to claim 2, wherein the exponential distribution law includes, as a parameter, a normalization factor with respect to at least one dimension of the images in the image sequence.

4. Method (2000) according to any one of claims 2 to 3, wherein the distribution law includes, as a parameter, an optimization factor, the optimization factor being previously defined by minimizing the difference between the community coefficient and a Jaccard index for non-occulted moving objects.

5. Method (2000) according to any one of claims 1 to 4, wherein the moving object is an individual in a substantially upright position and the surface distribution law G is a vertically decreasing distribution from the head to the feet of upright persons.

6. Method (2000) according to any one of claims 1 to 4, wherein the moving object is a vehicle and the surface distribution law G is a horizontally decreasing distribution in the direction of vehicle traffic.

7. Data processing device (4000) comprising means for implementing method (2000) according to any one of claims 1 to 6.

8. A computer program (14003) comprising instructions which, when the program is executed by a computer, cause the computer to implement the method according to any one of claims 1 QA

9. 1 dO. Computer-readable recording medium (4003) comprising instructions which, when executed by a computer, cause the computer to implement the method according to any one of claims 1 to 6.

10. A system for tracking a moving object in a sequence of images, said system comprises: - an image acquisition device; - a data processing device (4000) according to claim 7, said data processing device (4000) being configured to receive and process a sequence of images acquired by the image acquisition device.

Images