METHOD, DEVICE AND SYSTEM FOR REAL-TIME AND MULTI-CAMERA TRACKING OF A TARGET OBJECT

Description

[0001] The present invention relates to a method for real-time, multi-camera tracking of a target object. It also relates to a computer program, a device, and a system implementing such a method.

[0002] The field of the invention is generally the field of tracking objects from images captured by cameras. State of the art

[0003] Cities are increasingly equipping themselves with CCTV cameras, the number of which is growing faster than the number of human operators. It is very difficult to track a target object, such as a person or vehicle, as it moves from camera to camera. If the operator is disturbed, they may lose track of the target object, and finding it again can be particularly tedious. Tracking multiple target objects simultaneously is even more difficult. Tracking a target object across images provided by several cameras is also called multi-camera tracking.

[0004] We are familiar with multi-camera tracking solutions based on deep learning models, particularly re-identification models. These solutions suffer from a significant drop in performance on real-world data, i.e., images in actual use. This phenomenon is referred to in the state of the art as "domain drift." This performance drop results in a high rate of false positives and false negatives in object recognition, requiring a very large number of interactions with the operator to validate or invalidate the recognition suggestions made by the tracking solution, thus necessitating even more human operators.

[0005] US patent 2022 / 004768 A1 (PELLETIER BENOÎT [FR] ET AL) dated January 6, 2022, discloses a method to assist real-time monitoring of people in a sequence of images, using an analysis of tracklets previously calculated on targets identified in the images.

[0006] US patent 2021 / 182532 A1 (ARNAISE NICOLAS [FR] ET AL) dated June 17, 2021, discloses a tracking method to follow a reference element in successive images of an environment, the reference element being associated with a reference list of visual signatures, a visual signature being a vector describing the visual appearance of the reference element.

[0007] One object of the present invention is to remedy at least one of the drawbacks of the prior art.

[0008] Another aim of the invention is to provide a more efficient multi-camera tracking solution.

[0009] Another aim of the invention is to provide a multi-camera tracking solution that reduces interaction with human operators. Description of the invention

[0010] The invention proposes to achieve at least one of the aforementioned goals by a computer-implemented, real-time, multi-camera tracking method of a target object according to claim 1.

[0011] Thus, the invention proposes to determine, and in particular to adjust, the value of the human threshold based on the number of reference tracklets used to track the target object. For example, it is possible to decrease the value of the human threshold as the number of reference tracklets in the reference list increases. Alternatively, or in addition, it is possible to increase the value of the human threshold as the number of reference tracklets is small. In general, the value of the human threshold is adjusted inversely proportionally to the number of reference tracklets in the reference list.

[0012] Using a larger number of reference tracklets lowers the threshold at which operator validation is requested. Lowering this threshold reduces the number of times the human operator is prompted while tracking the target object, thus minimizing interaction with the operator.

[0013] Furthermore, using a larger number of reference tracklets increases tracking performance by reducing the number of false positives and false negatives.

[0014] In this application, "object" or "target object" means any type of object, such as a human, an animal, a car, etc.

[0015] In this application, "tracklet" means a set of at least one image, or image area, belonging to the same object and captured by a camera.

[0016] In this application, "tracklet passed" means a tracklet memorized by a camera prior to the tracking phase.

[0017] In this application, "new tracklet" means a tracklet stored by a camera during the tracking phase, i.e., after the tracking phase has been triggered.

[0018] The distance between two tracklets can be a Euclidean distance or a cosine distance.

[0019] The distance between two tracklets can be calculated using any known technique. Following a first example, and in no way limiting the possibilities, the distance between two tracklets can be determined based on the individual distances between each image forming one tracklet and each image forming the other tracklet, and then by calculating an aggregate distance based on these individual distances, for example, by averaging the individual distances. Alternatively, for a tracklet, it is possible to calculate an aggregate signature, for example, as the average of the signatures of each image composing that tracklet. The distance between two tracklets can then correspond to the distance between the aggregate signatures of the two tracklets.

[0020] As is known, the distance between two images can be calculated by generating a digital signature for each of the images, for example by an intelligent model such as a neural network, and then calculating the distance between these two digital signatures.

[0021] By "camera" we mean any type of image acquisition device, such as any type of RGB, LIDAR, thermal, 3D camera, etc.

[0022] At least one reference tracklet found in the reference list at the start of the process can be selected by the operator.

[0023] Alternatively, or in addition, at least one reference tracklet from the reference list at process start can be automatically selected / detected without operator intervention. For example, at least one reference tracklet from the reference list at process start can be detected following an alarm. In this case, the target object, and therefore the reference tracklet, can be automatically identified, for example, in the data associated with the alarm, thus enabling automatic tracking without operator intervention.

[0024] According to some embodiments, when the smallest first calculated distance is less than the re-identification threshold, the tracking phase may include an automatic assignment of the new tracklet to the target object.

[0025] Indeed, in this case, the object on the new tracklet is deemed sufficiently similar to the target object, so that the object on said new tracklet is considered to be the target object, without having to solicit the operator.

[0026] According to a particularly advantageous feature, when a new tracklet is assigned to the target object during the tracking phase, said tracking phase may further include a step of adding said new tracklet as a reference tracklet in said reference list.

[0027] Thus, the invention makes it possible to increase the number of reference tracklets in the reference list, which decreases the value of the human threshold and makes it possible to reduce the interaction with, and the solicitation of, the operator.

[0028] Assigning the new tracklet to the target object can be done automatically or manually, as described above.

[0029] According to an optional but particularly advantageous feature of the present invention, the method according to the invention may include, before the tracking phase, a preliminary phase of enriching the reference list, by adding at least one additional reference tracklet to said reference list, chosen from a database of past tracklets acquired during a predetermined period preceding the triggering of the tracking phase.

[0030] For example, the operator can choose a reference tracklet corresponding to the target object being tracked. This reference tracklet chosen by the operator can be used to identify other reference tracklets corresponding to the object and acquired during a past period immediately preceding the triggering of the target object's tracking.

[0031] Therefore, it is possible to increase the number of reference tracklets before starting to track the target object. As mentioned above, increasing the number of reference tracklets improves tracking performance and reduces the human threshold, thus minimizing operator interaction during tracking.

[0032] The predetermined period can be 30 minutes or one hour, immediately preceding the monitoring phase. This period can be adjusted by the operator.

[0033] Depending on the embodiment, the preliminary phase of enriching the reference list may include at least one iteration of the following steps performed for a past tracklet: calculation of at least one distance between: ▪ a reference tracklet from said reference list, and ▪ said passed tracklet, addition, or not, of said tracklet passed as additional reference tracklet in said reference list depending on: ▪ said at least one calculated distance, and ▪ the re-identification threshold and / or the human threshold.

[0034] Thus, the invention proposes to identify new reference tracklets among past tracklets as a function of the distance between the reference tracklet(s) and each past tracklet.

[0035] According to embodiments, for at least one tracklet passed, when at least one calculated distance is less than the re-identification threshold then the passed tracklet is added as a reference tracklet in the reference list.

[0036] Alternatively, or in addition, for at least one tracklet passed, when the smallest calculated distance falls between the re-identification threshold and the human threshold, the passed tracklet is presented to the operator. If the operator validates the passed tracklet as belonging to the target object, then said tracklet is added as a reference tracklet in the reference list. Otherwise, the passed tracklet is ignored.

[0037] According to an advantageous feature, the method according to the invention may include an update of the human threshold value in the event of the addition of a reference tracklet in said reference list.

[0038] For example, this update step can be performed each time a reference tracklet is added to the reference list, automatically and / or manually after validation by the operator, during the tracking phase and / or during the preliminary phase.

[0039] The value of the human threshold can be calculated using the following relationship: SH = SH i − NBTR × PA with: SH the value of the human threshold, SH i the initial value of the human threshold, chosen when starting the process according to the invention NBTR the number of reference tracklets in the reference list, PA a predefined learning step.

[0040] The PA learning step can be defined based on a maximum number, NB MAX, of reference tracklets in the reference list.

[0041] Alternatively, or in addition, the learning step size PA can be defined based on a total learning value, denoted D. For example: D ≤ SH i − SR , with SR being the re-identification threshold.

[0042] Following a non-limiting example implementation, the PA learning step size can be calculated according to the following relationship: PA = D / NB MAX

[0043] Following a non-exhaustive example of implementation: D = SH i − SR . In this case, the value of the human threshold can be calculated using the following relationship: SH n = SH i − NBTR × SH i − SR NB max In this case, as the number of reference tracklets (NBTR) in the reference list increases, the human threshold (SH) gradually approaches the re-identification threshold (SR), thus reducing the operator's workload during the tracking phase. Most importantly, when the number of reference tracklets (NBTR) in the reference list equals the maximum number of reference tracklets (NB max), then SH = SR, meaning that re-identification occurs fully automatically without operator intervention.

[0044] According to advantageous embodiments, the method according to the invention may further include an adjustment of the human threshold, and / or the re-identification threshold, as a function of at least one characteristic of a scene observed by a camera.

[0045] One such characteristic could be the scene's brightness. For example, it's possible to determine a scene's brightness based on the brightness of a tracklet or the brightness of images captured by the camera observing that scene. Specifically, the human threshold can be adjusted downwards as the scene's brightness increases and upwards as the scene's brightness decreases.

[0046] Such a characteristic could be, for example, the number of objects in the scene, and more generally, the scene's complexity. For instance, it is possible to determine the number of objects in a scene by counting them in an image captured by a camera observing that scene. Object detection can be performed using any known technique, such as the YOLOV5 algorithm. In particular, the human threshold can be adjusted downwards as the number of objects in the scene decreases and upwards as the number of objects in the scene increases.

[0047] According to an advantageous feature, the step of presenting a tracklet to the operator can be carried out according to at least one predetermined rule based on at least one of the following parameters: number of tracklets, noted NTP, to be presented to the operator at one time, elapsed time, DUR, since the last iteration of the presentation step, and / or number, NMT, of best matches.

[0048] In particular, the number of tracklets to be presented can correspond to the total number of tracklets to be presented to the operator. In this case, each tracklet that meets the human threshold (but not the re-identification threshold) is stored in a list of tracklets to be presented to the operator. As soon as the number of tracklets in this list reaches the NTP (Net Working Threshold), the presentation step can be performed. In one example, NTP = 10.

[0049] Alternatively, the NTP number can correspond to the maximum number of tracklets to be presented to the operator. In this case, even if the NTP number is not reached, for example due to a low number of new tracklets, then the tracklets can be presented to the operator when the DUR duration is reached.

[0050] In particular, the elapsed time (DUR) since the last iteration can correspond to the minimum time to be respected between two iterations of the presentation step. Following an example implementation, DUR = 10 minutes.

[0051] The number of best matches (NMT) can be the number of tracklets with the lowest first distance relative to the reference tracklets. In this case, only these tracklets are presented to the operator during the presentation step. Following a non-limiting example implementation, NMT = 3.

[0052] For example, when NTP=10 and NMT=3, the tracking phase is performed without the presentation step until there are at least 10 tracklets in the list to be presented to the operator. As soon as the list contains 10 tracklets, the presentation step is performed. During this presentation step, only the three tracklets with the best scores—that is, the lowest first distances relative to the reference tracklets—are presented to the operator. The other tracklets are ignored.

[0053] According to some embodiments, the validation, or not, of the operator for a first tracklet can be used to validate or not a second tracklet following said first tracklet, so that said second tracklet is not presented to the operator.

[0054] For example, the first and second tracklets can be obtained by the same camera in a very short time lapse.

[0055] The first and second tracklets can be obtained by splitting a source tracklet into several fragments due to occlusion, for example. In this case, the two fragments may be very similar, and the user's response to one of the fragments can likely be used for the other.

[0056] When the user has not yet been asked about the first tracklet, or has not yet taken a position on the first tracklet, the second tracklet can be placed in a queue pending validation or rejection of the first tracklet. Then, once the operator has made a decision regarding the first tracklet—validating or rejecting it—the same decision is applied to the second tracklet.

[0057] The follow-up phase may also include a step of memorizing the new tracklet in a list, called the rejected tracklet list, when said new tracklet is not assigned to the target object.

[0058] The new tracklet can be rejected when the smallest first distance is greater than the human threshold.

[0059] The new tracklet can be rejected when the smallest first distance is between the human threshold and the re-identification threshold but is rejected following a decision by the operator during the interrogation step.

[0060] The list of rejected tracklets can include all rejected tracklets. Alternatively, the list of rejected tracklets can include only a predetermined number of the most recently rejected tracklets. As yet another alternative, the list of rejected tracklets can include only tracklets rejected during a time window, for example, 15 minutes, preceding the current time.

[0061] According to an advantageous feature, the monitoring phase may also include, before the step of presenting the new tracklet to the operator, a testing step comprising the following steps: calculation of at least one second distance between the new tracklet and at least one previously rejected tracklet, and rejection of said new tracklet as not belonging to the target object when: ▪ the smallest first distance is greater than the smallest second, and ▪ the smallest second distance is less than a threshold, called the similarity threshold, predetermined;

[0062] Thus, the method according to the invention makes it possible to automatically reject a new tracklet, as not belonging to the target object, when said new tracklet presents a significant resemblance to at least one previously rejected tracklet. Therefore, this new tracklet is not presented to the operator, thereby reducing the operator's workload.

[0063] A new tracklet thus rejected can be stored in the list of rejected tracklets.

[0064] The similarity threshold can be of a fixed value. The similarity threshold can be set at the start of the process according to the invention.

[0065] It can happen that the operator rejects too many tracklets. This can be a sign that the human threshold value (HS) is too high and that too many false positives are being presented to the operator. In this case, the method according to the invention may include a readjustment step, and in particular a decrease, of the human threshold value by a predetermined readjustment step.

[0066] The readjustment step can be the same as, or different from, the learning step described above.

[0067] The readjustment step can be a fixed value, or a variable value, for example depending on a number of tracklets rejected by the operator.

[0068] The human threshold value readjustment step SH can for example be triggered when the number of tracklets rejected over a predetermined time window, in particular sliding, reaches a predetermined threshold value.

[0069] According to another aspect of the invention, a computer program is proposed comprising executable instructions which, when executed by a computer device, implement all the steps of the method, according to the invention, of real-time and multi-camera tracking of a target object.

[0070] The computer program can be in any computer language, such as for example machine language, C, C++, JAVA, Python, etc.

[0071] According to another aspect of the invention, a real-time, multi-camera tracking device for a target object is proposed, comprising means configured to implement all the steps of the method, according to the invention, for real-time, multi-camera tracking of a target object.

[0072] The device according to the invention can be any type of device such as a server, a computer, a tablet, a calculator, a processor, a computer chip, programmed to implement the method according to the invention, for example by the computer program according to the invention.

[0073] According to another aspect of the invention, a real-time, multi-camera tracking system for a target object is proposed, comprising: several cameras, and a device according to the invention for real-time and multi-camera tracking of a target object from images provided by said cameras.

[0074] At least one of the cameras can be local or remote from the device. In the latter case, the camera can communicate with the device via a wired or wireless communication network, such as the Internet.

[0075] Cameras can be distributed in space, for example within a place such as a train station or airport, or in a city, or along a traffic route. Description of the figures and methods of implementation

[0076] Other advantages and features will become apparent upon examination of the detailed description of non-limiting embodiments and the accompanying drawings, in which: there FIGURE 1 is a schematic representation of a first, non-limiting example of a method according to the invention; the FIGURE 2 is a schematic representation of a second, non-limiting embodiment of a method according to the invention; the FIGURE 3is a schematic representation of a third, non-limiting embodiment of a method according to the invention; the FIGURE 4 is a schematic representation of a non-limiting example embodiment of a device according to the invention; and the FIGURE 5 is a schematic representation of a non-limiting example embodiment of a system according to the invention.

[0077] It is understood that the embodiments described below are by no means exhaustive. In particular, variants of the invention may be conceived comprising only a selection of the features described below, isolated from the other features described, if this selection of features is sufficient to confer a technical advantage or to differentiate the invention from the prior art. This selection includes at least one preferably functional feature without structural details, or with only a portion of the structural details if this portion alone is sufficient to confer a technical advantage or to differentiate the invention from the prior art.

[0078] In particular, all the variants and embodiments described can be combined with each other if there are no technical obstacles to this combination.

[0079] In the figures and in the rest of the description, elements common to several figures retain the same reference.

[0080] There FIGURE 1 is a schematic representation of a first non-limiting example of an embodiment of a method according to the present invention.

[0081] The 100 process of the FIGURE 1 can be used for tracking any type of target object such as humans, animals, cars, etc. from images from multiple cameras distributed in space.

[0082] In the following, and without loss of generality, it is assumed that the process is used for tracking a target person.

[0083] The method 100 includes a step 102 in which an operator selects at least one reference tracklet corresponding to the target person to be tracked. Preferably, but not exclusively, each reference tracklet is obtained from images from one of the cameras captured prior to tracking. Each reference tracklet is stored in a list of reference tracklets, also called the reference list.

[0084] Process 100 then includes a phase 104 of tracking the target person from images from the cameras.

[0085] During tracking phase 104, each new tracklet from each camera, corresponding to a person, is analyzed to determine whether or not the new tracklet matches the target person. To do this, tracking phase 104 includes step 106, in which a cosine or Euclidean distance, called the first distance, is determined between each reference tracklet and the new tracklet. Thus, in step 106, for the new tracklet, as many first distances are determined as there are reference tracklets in the reference list. Of these first distances, only the smallest, denoted PDI min, is retained. Alternatively, the average of the calculated first distances can be retained.

[0086] During step 108, the smallest calculated first distance PDI min 108 is compared to a predetermined re-identification threshold, SR. For example, SR=0.1.

[0087] If PDI min ≤SR, then the new tracklet is automatically assigned to the target person. During step 110, the new tracklet is added to the reference list as a reference tracklet.

[0088] If PDI min > SR, then, during step 112, the PDI min distance is compared to a predetermined human threshold, SH, with SH > SR. For example, SH = 0.4. If PDI min > SH, then the new tracklet is rejected as not belonging to the target object.

[0089] However, if in step 112, PDI min ≤ SH, then the new tracklet can be presented to the operator for validation in step 114. If the operator does not validate the new tracklet as belonging to the target person, then the new tracklet is rejected as not belonging to the target object. If, in step 114, the operator validates the new tracklet as belonging to the target person, then the new tracklet is added to the reference list as the reference tracklet in step 110.

[0090] During step 114, the new tracklet can be presented to the operator individually. Alternatively, the new tracklet can be stored in a list of tracklets to be presented to the operator, called the validation list. New tracklets in the validation list can be presented to the operator based on at least one of the following parameters: number of tracklets, noted NTP, to be presented to the operator at one time, elapsed time, DUR, since the last iteration of the presentation step, and / or number, NMT, of best matches.

[0091] For example, when NTP=15 and NMT=3, then the follow-up phase 104 is performed without the presentation step 114 until there are at least 15 tracklets in the validation list. As soon as the validation list contains 15 tracklets, the presentation step 114 is performed. During this presentation step, only the three tracklets with the best scores—that is, the smallest distances from the reference tracklets—are presented to the operator. The other tracklets are ignored.

[0092] Advantageously, method 100 further includes a step 116 of updating the human threshold value each time a reference tracklet is added to the reference list. Following an example embodiment, and without loss of generality, the human threshold value SH can be updated according to the following relationship: SH = SH i − NBTR . SH i − SR NB max SH the value of the human threshold; SH i the initial value of the human threshold chosen when starting process 100; NBTR the number of reference tracklets in the reference list; NB max the maximum number of reference tracklets in the reference list; and PA a predefined learning step.

[0093] The PA learning step can be defined based on a maximum number, NB MAX, of reference tracklets in the reference list.

[0094] Alternatively, or in addition, the learning step size PA can be defined based on a total learning value, denoted D. For example: D ≤ SH i − SR .

[0095] Following a non-limiting example implementation, the PA learning step size can be calculated according to the following relationship: PA = D / NB MAX

[0096] Following a non-exhaustive example of implementation: D = SH i − SR .

[0097] Taking SR=0.1, SH i =0.4, as indicated above as a non-limiting example, we have: SH = 0.4 − NBTR . 0.3 NB max Taking as an example, NB max = 600 then SH = 0.4 − NBTR ⋅ 0.3 600 Thus, the more reference tracklets there are in the reference list, the more the value of the human threshold SH increases and approaches the value of the re-identification threshold, so that the operator is less solicited as the follow-up phase progresses.

[0098] When the new tracklet is rejected at step 112 because PDI min >SH, or when the new tracklet is rejected at step 114 because the operator indicates that the new tracklet does not belong to the target object, the follow-up phase may optionally include a step 118 of adding the new tracklet to a list of rejected tracklets.

[0099] Furthermore, in step 112, when PDI min ≤ SH, the follow-up phase 104 may include an optional testing step 120, performed before the presentation step 114. The purpose of this testing step 120 is to determine whether the new tracklet is similar to a previously rejected tracklet.

[0100] Test step 120 includes step 122, which calculates at least one distance, Euclidean or cosine, called the second distance, denoted DDI, between the new tracklet and each tracklet in the list of rejected tracklets. Thus, in step 122, for the new tracklet, as many second distances are determined as there are rejected tracklets in the list of rejected tracklets. Among these second distances, only the smallest second distance, denoted DDI min, is retained.

[0101] During stage 124: the smallest second distance DDI min is compared to the smallest first distance PDI min, and the smallest second distance DDI min is compared to a predetermined threshold, called the similarity threshold, denoted SS;

[0102] When, DDI min <PDI min Andthan DDI min <SS, alors on estime que la nouvelle tracklet est très similaire à une des tracklets déjà rejetées. Dans ce cas, la nouvelle tracklet est rejetée comme n'appartenant pas à l'objet cible. La nouvelle tracklet peut optionnellement être mémorisée lors de l'étape 118 dans la liste des tracklets rejetées. Dans le cas contraire, alors la phase de suivi 104 continue à l'étape 114, de manière similaire à ce qui a été décrit plus haut.

[0103] The similarity threshold SS can be a fixed value. The similarity threshold can be set at the start of the process according to the invention.

[0104] There FIGURE 2 is a schematic representation of a second, non-limiting example of an embodiment of a method according to the present invention.

[0105] The 200 process of the FIGURE 2This includes step 102, in which the operator selects at least one reference tracklet corresponding to the target person to be tracked. Preferably, each reference tracklet is obtained from images from one of the cameras captured prior to tracking. Each reference tracklet is stored in a list of reference tracklets, also called the reference list.

[0106] The method 200 then includes a phase 204 of enriching the reference list by adding at least one additional reference tracklet to said reference list, chosen from a database of past tracklets acquired during a predetermined period preceding the tracking phase. In other words, the enrichment phase 204 is carried out just before the execution of the tracking phase 104. The past tracklets are those that were acquired by the cameras during a predetermined past period immediately preceding the tracking phase 104. For example, the past period could be 30 minutes or one hour.

[0107] The enrichment phase 204 is carried out in turn and individually for each tracklet passed, corresponding to a person, and which was acquired by one of the cameras during the predetermined period immediately preceding the tracking phase.

[0108] The enrichment phase includes a step 206 which calculates, for the passed tracklet, a cosine or Euclidean distance, denoted DI, between said passed tracklet and each reference tracklet in the reference list. Of the calculated distances, only the smallest DI min is retained.

[0109] During step 208, the smallest calculated distance DI min during step 208 is compared to the predetermined re-identification threshold, SR. For example, SR=0.1.

[0110] If DI min ≤SR, then the passed tracklet is automatically assigned to the target person. During step 210, the passed tracklet is added to the list of reference tracklets as a reference tracklet, thus enriching the reference list.

[0111] If DI min > SR, then, at step 212, the distance DI min is compared to a predetermined human threshold, SH, with SH > SR. For example, SH = 0.4. If DI min > SH, then the passed tracklet is rejected.

[0112] However, if DI min ≤ SH, then the passed tracklet is presented to the operator for validation during step 214. If the operator does not validate the passed tracklet as belonging to the target person, then the passed tracklet is rejected, and the enrichment phase 204 can be performed for another passed tracklet, if necessary. If the operator validates the passed tracklet as belonging to the target person, then the passed tracklet is added as a reference tracklet to the reference list during step 210.

[0113] The enrichment phase 204 can be repeated for each past tracklet, as long as there are past tracklets acquired during the predetermined past period remaining.

[0114] Thus, the enrichment phase allows the reference list to be enriched with one or more passed tracklets, from one or more reference tracklets selected by the operator.

[0115] The process 200 includes, after the enrichment phase 204, a step 216 of updating the human threshold according to the number of reference tracklets stored in the reference list.

[0116] The update of the human threshold value can be performed according to any predetermined relationship.

[0117] Following a specific implementation example and without loss of generality, the value of the human threshold can be updated according to the following relationship: SH = SH i − NBTR . SH i − SR NB max SH the value of the human threshold; SH i the initial value of the human threshold; NBTR the number of reference tracklets in the reference list; NB max the maximum number of reference tracklets in the reference list; and PA a predefined learning step.

[0118] Step 216, which updates the SH human threshold value, can be performed each time a passed tracklet is added to the reference list during phase 204. Alternatively, step 216 can be performed only once, when all iterations of the enrichment phase have been completed and there are no more passed tracklets to process. As yet another alternative, step 216 can be performed at a different frequency, for example, one chosen by the operator.

[0119] Optionally, the enrichment phase 204 may include a step of memorizing the passed tracklet into a list of rejected passed tracklets when it is rejected.

[0120] In addition, optionally, the enrichment phase 204 may further include a test step, similar to the test step 120, carried out with the previously rejected past tracklets stored in the list of previously rejected past tracklets.

[0121] The process 200 may further include a follow-up phase 104, as described above with reference to the FIGURE 1 .

[0122] It should be noted that, unlike the 100 process of the FIGURE 1 , process 200 of the FIGURE 2 does not include step 116 of updating the SH human threshold value during monitoring.

[0123] There FIGURE 3 is a schematic representation of a third, non-limiting embodiment of a method according to the present invention.

[0124] The 300 process of the FIGURE 3 includes a combination of processes 100 and 200 of FIGURES 1 And 2. Process 300 includes step 102, enrichment phase 204, step 216 of updating the value of the human threshold SH before the start of monitoring, monitoring phase 104 and step 116 of updating the value of the human threshold SH during monitoring.

[0125] There FIGURE 4 is a schematic representation of a non-limiting example embodiment of a device according to the present invention.

[0126] The 400 device of the FIGURE 4 can be used to implement a method for tracking a target object according to the invention, and in particular any one of the methods 100, 200 or 300 described above.

[0127] Device 400 includes at least one module 402 for operator selection of at least one reference tracklet. This module allows the operator to add one or more reference tracklets to a reference list (LREF) at the beginning of the process. Specifically, module 402 is configured for the implementation of step 102 described above.

[0128] Device 400 includes at least one module 404 for enriching the LREF reference list from past tracklets stored in a database B1 of past tracklets. In particular, module 404 is configured to implement enrichment phase 204 of the FIGURES 2 Or 3 .

[0129] The 400 device includes at least one 406 tracking module enabling the tracking of a target object and, optionally, the enrichment of the LREF reference list each time a new tracklet corresponding to the target object is identified. In particular, the 406 module is configured to implement tracking phase 104 of the FIGURES 1 , 2 Or 3 .

[0130] Device 400 includes at least one module 408 for updating the human threshold value based on the number of reference tracklets in the reference list and a predetermined relationship. In particular, module 408 is configured to implement at least one, and specifically each, of update steps 116 and 216 of the FIGURES 1 , 2 Or 3 .

[0131] At least one of the 402-408 modules can be independent of the other 402-408 modules. At least two of the 402-408 modules can be integrated within the same module.

[0132] Each 402-408 module can be a hardware module or a software module, such as a computer application or program, executed by an electronic component such as a processor, electronic chip, computer, etc.

[0133] There FIGURE 5 is a schematic representation of a non-limiting example embodiment of a tracking system according to the invention.

[0134] System 500 can be used to implement a tracking method according to the invention, and in particular any one of the methods 100, 200 or 300 of FIGURES 1 , 2 Or 3 .

[0135] The 500 system is used for tracking a target object of any type, and in particular a target person, from video streams from K 502 1 -502 K cameras of said system, distributed in space and forming a 504 network of cameras, with K≥1. In the example shown, and in no way limiting, K=3.

[0136] The 500 system further includes a tracking device according to the invention, such as, for example, the 400 device of the FIGURE 4 .

[0137] Device 400 can be local to cameras 502 1 -502 3 or remote from cameras 402 1 -402 3.

[0138] The 400 device can be connected to each of the 502 1 -502 3 cameras via a wired or wireless connection.

[0139] Of course, the invention is not limited to the examples just described. For example, the number of cameras is not limited to 3. Furthermore, the threshold values ​​and the number of reference tracklets, as well as the various relationships mentioned for calculating the human threshold, are given as a non-limiting example.

Claims

1. A computer implemented method (100;200;300) for real-time multi-camera tracking of a target object, from a list (LREF), called reference list, comprising at least one tracklet, called reference tracklet, belonging to said target object, said method (100;200;300) comprising a tracking phase (104) carried out for at least one new tracklet captured by at least one of the cameras (5021-5023), said tracking phase comprising the following steps: - calculating (106) at least a first distance between said new tracklet and at least one reference tracklet; - when the smallest calculated first distance is between a first predetermined threshold (SR), called re-identification threshold, and a second predetermined threshold (SH), called human threshold, greater than said re-identification threshold, presenting (114) said new tracklet to an operator, for said operator to assign or not assign said new tracklet to said target object, manually; characterized in that the human threshold (SH) is a function of the number of reference tracklets in said reference list (LREF), and in that the human threshold value is adjusted in a manner inversely proportional to the number of reference tracklets in the reference list.

2. The method (100;200;300) according to the preceding claim, characterized in that, when the smallest calculated first distance is less than the re-identification threshold (SR), the tracking phase (104) comprises an automatic assignment of the new tracklet to the target object.

3. The method (100;200;300) according to any one of the preceding claims, characterized in that, when a new tracklet is assigned to the target object during the tracking phase, said tracking phase (104) further comprises a step (110) of adding said new tracklet as a reference tracklet in said reference list (LREF).

4. The method (200;300) according to the preceding claim, characterized in that it comprises, before the tracking phase (104), a prior phase (204) of enriching the reference list (LREF), by adding at least one additional reference tracklet in said reference list, chosen from a database (B1) of past tracklets acquired during a predetermined period preceding the triggering of the tracking phase (104).

5. The method (200;300) according to the preceding claim, characterized in that the prior phase (204) of enriching the reference list (LRFE) comprises at least one iteration of the following steps carried out for a past tracklet: - calculating (206) at least one distance between: ∘ a reference tracklet of said reference list, and ∘ said past tracklet, - adding (210), or not, said past tracklet as an additional reference tracklet in said reference list (LREF) as a function of: ∘ said at least one calculated distance, and ∘ the re-identification threshold (SR) and / or the human threshold (SH).

6. The method (100;200;300) according to any one of claims 3 to 5, characterized in that it comprises an update (116;216) of the value of the human threshold (SH) in the event a reference tracklet is added to said reference list (LREF).

7. The method (100;200;300) according to any one of the preceding claims, characterized in that it comprises an adjustment of the human threshold (SH), and / or of the re-identification threshold (SR), as a function of at least one characteristic of a scene observed by a camera (5021-5023), said characteristic being the brightness of the scene or the complexity of the scene.

8. The method (100;200;300) according to any one of the preceding claims, characterized in that the step (114) of presenting a tracklet to the operator is carried out according to at least one predetermined rule as a function of at least one of the following parameters: - number of tracklets to be presented to the operator at one time, - elapsed time since the last iteration of the presentation step, and / or - number of best matches.

9. The method (100;200;300) according to any one of the preceding claims, characterized in that the same decision by the operator for a first tracklet, validation or rejection of the first tracklet, is applied to a second tracklet following said first tracklet, validation or rejection of the second tracklet.

10. The method (100;200;300) according to any one of the preceding claims, characterized in that it comprises a step (118) of storing the new tracklet in a list, called rejected tracklet list, when said new tracklet is not assigned to the target object.

11. The method (100; 200; 300) according to any one of the preceding claims, characterized in that the tracking phase (104) further comprises, prior to the step (114) of presenting the new tracklet to the operator, a test step (120) comprising the following steps: - calculating (122) at least one second distance between the new tracklet and at least one previously rejected tracklet, and - rejecting said new tracklet as not belonging to the target object when: ∘ the smallest first distance is greater than the smallest second distance, and ∘ the smallest second distance is less than a predetermined threshold, referred to as the similarity threshold (SS);12. A computer program comprising executable instructions which, when executed by a computing apparatus, implement all the steps of the method (100;200;300) for real-time multi-camera tracking of a target object according to any one of the preceding claims.

13. A device (400) for real-time multi-camera tracking of a target object comprising means configured to implement all the steps of the method (100;200;300) for real-time multi-camera tracking of a target object according to any one of claims 1 to 11.

14. A system (500) for real-time multi-camera tracking of a target object, comprising: - a plurality of cameras (5021-5023), and - a device (400) according to the preceding claim for real-time multi-camera tracking of a target object from images provided by said cameras.