METHOD AND DEVICE FOR CALIBRINGING A COLLABORATIVE TARGET TRACKING SYSTEM

Description

TECHNICAL FIELD OF THE INVENTION

[0001] The present invention relates to the field of detection and more particularly the tracking of targets or objects. STATE OF PRIOR ART

[0002] Multi-target tracking is a complex task requiring the implementation of several algorithms generating a large number of operations. Multi-target tracking involves the simultaneous use of several hardware components: a standard processor, specialized video decoding chips, and graphics processors.

[0003] The algorithms used by the processing chain of a multi-target tracking system include a significant number of configuration parameters influencing both the quality of the tracking obtained and the load imposed on the hardware.

[0004] The system must first characterize all known targets and then assess whether a detection is compatible with a target. The characterization can focus, for example, on movement or visual appearance. The system must therefore establish a correlation metric between a detection and a target and choose the pair (target, detection) with the best metric.

[0005] When the correlation metric is unsatisfactory for a set of known targets, the tracking system must establish that a new target has been perceived. Thus, the system must rely on a threshold for the correlation metric. This threshold depends on the technologies and methods used for target characterization and detection, but also on the class of targets considered. The class of a target allows the nature of the target to be defined. For example, and without limitation, "Human", "car", "dog" are classes that can be processed and tracked by the system.

[0006] The situations for which a tracking or target detection algorithm of a tracking system does not work can be used by said tracking system to prevent said tracking algorithm, when executed on the processor of a tracking device operating in its target environment, from generating erroneous detection information. Determining the situations for which a tracking algorithm does not work generally consists of reproducing the operating conditions of the tracking algorithm in the evaluation phase and comparing the results of said algorithm with a ground truth. However, this approach poses difficulties ranging from determining the situations that the tracker will be confronted with to repeating the evaluation each time a new target environment arises and / or the target environment is likely to change. STATEMENT OF THE INVENTION

[0007] The present invention aims to overcome certain drawbacks of the prior art by providing a method for improving target tracking or detection in a given environment.

[0008] For this purpose, the present invention relates to a method for calibrating a single or multi-target collaborative tracking system comprising at least one set of tracking algorithms T executed by at least one processor to process information from an image stream for the detection of targets in the video image stream captured at least by a given camera of the tracking system, according to a given viewpoint of a scene to be monitored, each tracking algorithm T using at least one known reference area S ref T< representing all the situations for which the algorithm detects a real target and at least one known failure area S fail T< representing all the situations for which the tracking algorithm does not detect a real target,the single or multi-target tracking system comprising at least one set of modules / programs recorded in a memory and at least one processor executing all of the modules to implement the calibration method, said method being characterized in that it comprises at least the steps consisting of: , Collect (E1) the associations of all tracking algorithms T of the single or multi-target tracking system to create a dictionary of predictions Dict predictions for a given target; Determine (E2), via the created prediction dictionary Dict predictions, T tracking algorithms in the reference area to create a “ground truth” list Ground truth listincluding the tracking algorithms T having a confidence score higher than a predefined threshold value; Update (E3) the known failure zone of each tracking algorithm T in the list of tracking algorithms T of the single or multi-target tracking system not part of the “ground truth” list Ground truth list and having a likelihood score below a predetermined threshold.

[0009] According to another feature, the step of collecting (E1) the associations of tracking algorithms consists of at least

[0010] For each tracking algorithm T: Predict (E10), from the video image stream, the states of a target or a set of targets contained in the image streams to create a dictionary of predictions Dict predictions T< linked to the tracking algorithm T, said dictionary of predictions Dict predictions T< taking as input an existing or new target ID Target IDand giving as output value a vector characterizing the state of the new target named Target state; For each triplet (Target ID, Target State, Date) included in the prediction dictionary Dict predictions T< of the tracking algorithm T, create (E11) the prediction dictionary Dict predictions for a given target by adding to a list the triplet (tracker ID, target state, date), the prediction dictionary Dict predictions for a given target taking as input the target identifier IDcible and giving as output value a list of triples (ID tracker, Target state, date) where ID tracker is the identifier of the tracking algorithm T and the date is the value of the date on which the detection of the target was carried out by the tracking algorithm T;

[0011] According to another feature, the step of determining (E2) tracking algorithms T in the reference zone comprises

[0012] For each target Target ID and triple list (Tracker ID, Target State, Date) in a target's prediction dictionary Dict predictions Calculate (E20) a vector of metrics Target site of the target, the metric vector characterizing the situation of the target Target ID; Check if the metric vector Target site belongs to the reference area S ref T< of the tracking algorithm T; If yes, add the quintuplet (target ID, bestTracker ID, target State, target Sit, date) to a list to create (E21) the “ground truths” list List of ground truths, ID bestTracker designating the tracking algorithm T of the triplet list included in the prediction dictionary Dict predictions and having the highest confidence score, in the calculation of the situation characterized by the vector of metrics Target site, in comparison with other scores obtained by other algorithms in said prediction dictionary

[0013] According to another particularity, the update step (E3) includes: For each element (target ID, bestTracker ID, target state GT<, target Sit, date) of the “ground truths” list Ground truth list at least one of the steps consisting of: Search (E30), in the prediction dictionary for a given target Dict predictions, the prediction for the tracking algorithm T with the identifier ID tracker and the target ID target at the given date, the prediction corresponding to the target state Target state in the list (ID tracker , Target state , date) of the prediction dictionary Dict predictions, Calculate (E31) a likelihood score defined by measuring the distance between the target state and the “ground truth” list List of ground truths, Target state GT< , and the target state searched in the list (tracker ID , Target state , date) of the prediction dictionary Dict predictions If the identity ID tracker of the tracking algorithm T of the prediction dictionary Dict predictionsis different from the bestTracker ID identity of the tracking algorithm T from the "ground truths" list Ground truth list and that the calculated likelihood score is less than a predetermined threshold value, then; Update (E32) the failure zone S fail T< of the tracking algorithm T by adding the metrics vector Target site calculated in the step of determining the tracking algorithm T in the reference area and included in the element element (target ID, bestTracker ID, target state GT<, target Sit, date) of the “ground truths” list Ground truth list

[0014] According to another aim, the invention also relates to a device making it possible to improve the tracking or detection of a target in a given environment.

[0015] This aim is achieved by a calibration device (1) of a single or multi-target tracking system comprising at least one memory for storing a set of programs and at least one calculation unit for executing said programs to carry out the steps of the calibration method according to any one of the preceding claims and to store the results of the calculations in the lists, (List of ground truths, List of the dictionary of predictions Dict predictions containing each triplet (tracker ID, target state, date) of a target target ID), the prediction dictionaries, the reference area S ref T< or the failure area S fail T< known to the tracking algorithm T for each tracking device (tracker) of a plurality of tracking devices of the tracking system, each tracking device comprising at least one tracking algorithm T..

[0016] According to another feature, the calibration method is used in a single or multi-target collaborative tracking system in which the image stream from a video stream server (SFV) is sent, on the one hand to an object detection device (RFCN) to deliver information for identifying new targets to each of the single-target tracking devices DT MIL and DT KCF executing tracking algorithms of the MIL (Multiple Instance Learning) and KCF (kernelized correlation filter) type, on the other hand to each of the tracking devices DT MIL and DT KCF whose output signals are sent for processing by the calibration device (1), implementing the calibration method, and comprising a device (DRCTS) executing a main collaborative tracking routine and dialoguing on the one hand with a situation calculator (CS) to obtain the metrics vector Target sitecalculated for each new target and with a manager (GZE) of the failure zones S fail T< known to provide this information to a main tracking device (DTP) developing the tracking result from the information transmitted by GZE and the information transmitted by the tracking device DT MIL or DT KCF.: According to another particularity, the calibration method is used in a single or multi-target collaborative tracking system in which the image stream from a video stream server is sent, on the one hand to a multi-target tracking device DT MHT executing a tracking algorithm of the MHT (Multiple Hypothesis Tracking) type and an object detection device (RFCN) to deliver information identifying new targets to said multi-target tracking device DT MHT,on the other hand to each of two sets of a plurality of tracking devices DT MIL and DT KCF executing tracking algorithms of the MIL (Multiple Instance Learning) and KCF (kernelized correlation filter) type, the multi-target tracking device DMHT transmitting target initialization information comprising at least the coordinates of said targets to each of the two sets of a plurality of tracking devices DT MIL and DT KCF, the output signals of each of the tracking devices DT MHT, DT MIL, DT KCF are sent for processing by the calibration device (1), according to claim 5, implementing the calibration method, according to claims 1 to 4, and comprising a device (DRCTS) executing the main collaborative tracking routine (RCTS) and dialoguing on the one hand with a situation calculator (CS) (Sit) to obtain the metrics vector, Target sitecalculated for each new target and with a manager (GZE) of the known failure zones S fail T< to provide this information to a main tracking device (DTP) developing the tracking result from the information transmitted by GZE and the information transmitted by each tracking device DT MHT, DT MIL or DT KCF.

[0017] Other features and advantages of the present invention are detailed in the following description. BRIEF DESCRIPTION OF THE FIGURES

[0018] Other features and advantages of the present invention will appear more clearly on reading the following description, given with reference to the appended drawings, in which: [ Fig 1A ] - There Figure 1A schematically represents the steps of the method for calibrating the failure zones associated with the tracking algorithms T of a single or multi-target tracking system, according to one embodiment; [ Fig 1B ] - There Figure 1Bschematically represents the step of collecting (E1) the associations of all the tracking algorithms T of the single or multi-target tracking system, according to one embodiment; [ Fig 1C ] - There Figure 1C schematically represents the step of determining (E2) the tracking algorithms T, of the single or multi-target tracking system, which are in the reference zone, according to one embodiment; [ Fig 1D ] - There Figure 1D schematically represents the step of updating (E3) the failure zone of each tracking algorithm T of the single or multi-target tracking system, according to one embodiment; [ Fig 2 ] - There figure 2 schematically represents the integration of the calibration device (1) in the processing chain of a single or multi-target tracking or detection system, according to one embodiment; [ Fig 3 ] - There figure 3schematically represents the use of the calibration device (1) in the processing chain of a single-target tracking or detection system, according to one embodiment; [ Fig 4 ] - There figure 4 schematically represents the use of the calibration device (1) in the processing chain of a multi-target tracking or detection system, according to one embodiment; DESCRIPTION OF EMBODIMENTS

[0019] The present invention relates to a method for calibrating a single or multi-target collaborative tracking system.

[0020] In some embodiments, the calibration method is implemented in a single- or multi-target collaborative tracking system comprising at least one set of tracking algorithms T executed by at least one processor to process information from an image stream for detecting targets in the video image stream captured at least by a given camera of the tracking system, according to a given viewpoint of a scene to be monitored. Each tracking algorithm T uses at least one known reference area S ref T< which represents all the situations for which the algorithm detects a real target and at least one known failure area S fail T< representing all the situations for which the tracking algorithm does not detect a real target.

[0021] The situation or situation of a target named "cx" is characterized by a metrics vector called Sit cx. The metrics vector Sit cx may include, for example and in a non-limiting manner, the following components: Coordinates in the camera plane Target size "cx" Target speed "cx" Number of neighboring targets Difference between the dominant color of the target and the dominant color of the background.

[0022] The metrics vector thus provides information that can be used to identify a target.

[0023] The tracking system comprises at least one memory in which each tracking algorithm T records, in a dedicated space, its success situations (target detection) recovered during tests on training data from video image streams stored in a database of said system. These success situations constitute a reference zone. “Detecting a real target” therefore corresponds to giving a prediction in one of the situations in the reference zone.

[0024] The reference zone S ref T< of a tracking algorithm T is formed by metric vectors for which the tracking algorithm gives, for example, and in a non-limiting manner, a confidence score greater than a predetermined threshold.

[0025] In a variant, the reference area S ref T< of a tracking algorithm T can be obtained by statistical observation on a large number of data.

[0026] In another variant, the reference area S ref T< of a tracking algorithm T can be obtained by evaluation (inference) of a set of conditions (rules), predefined and stored in the collaborative tracking system, characterizing a situation.

[0027] By calibrating the tracking system, we mean the determination of failure zones or a set of situations for which the tracking algorithms T of the tracking system do not detect targets or do not work. The single or multi-target tracking system comprises at least one set of modules / programs stored in a memory and at least one processor executing all the modules to implement the calibration method. In some embodiments, the calibration method comprises at least one step of collecting (E1) the associations of all the tracking algorithms T of the single or multi-target tracking system to create a dictionary of predictions Dict predictions for a given target; In a variant, the calibration method comprises at least one step consisting of determining (E2), via the dictionary of predictions created Dict predictions, T tracking algorithms in the reference area to create a “ground truth” list Ground truth list comprising the tracking algorithms T having a confidence score greater than a predefined threshold value; In another variant, the calibration method comprises at least one step consisting of updating (E3) the known failure zone of each tracking algorithm T of the list of tracking algorithms T of the single or multi-target tracking system not forming part of the “ground truth” list Ground truth list and having a likelihood score below a predetermined threshold.

[0028] In some embodiments, after receiving the video image streams captured by at least one camera of the single or multi-target collaborative tracking system such as, the step of collecting (E1) the tracking algorithm associations, as illustrated by the Figures 1A and 1B , consists of at least: For each tracking algorithm T: Predict (E10), from the video image stream, the states of a target or a set of targets contained in the image streams to create a dictionary of predictions Dict predictions T< related to the tracking algorithm T. Said prediction dictionary Dict predictions T< takes as input a target ID Target ID, assigned to an existing or new target, and outputs as a value a vector characterizing the state of the new target named Target State ; The Dictionary of Predictions Dict predictions T<consists of the predictions calculated by the tracking algorithm T, executed on a processor of the collaborative tracking system or on the processor of a tracking device of said collaborative tracking system and in which the tracking algorithm is stored, concerning an instant "t+1" from data (stream of video images recorded in a database of the collaborative tracking system) known at the instant "t". The tracking algorithm T proposes predictions by associating a detection with a target with a confidence score. The state of the new target Target state may include for example the position of the target in the image, the confidence score calculated by the tracking algorithm, etc.

[0029] For each triplet (Target ID, Target State, Date) included in the prediction dictionary Dict predictions T< of the tracking algorithm T, create (E11) the prediction dictionary Dict predictions for a given target by adding to a list the triplet ( ID tracker , Target state, date). The prediction dictionary Dict predictions for a given target takes as input the target ID Target ID and gives as output value a list of triples ( ID tracker, Target state, date) where ID tracker is the identifier assigned to the tracking algorithm T and “date” is the value of the date on which the target detection was performed by the tracking algorithm T. For each target, Dict predictions is the union or set of Dict predictions T< of each of the tracking algorithms T for this target.

[0030] In some embodiments, once the dictionaries Dict predictions targets were created, the step of determining (E2) tracking algorithms T in the reference area of ​​each target, as illustrated by the Figures 1A and 1C , is implemented. This determination step (E2) includes at least the following sub-steps: For each target Target ID and triple list ( ID tracer, Target state, date) in a target's prediction dictionary Dict predictions created in the collection step (E1) of tracking algorithm associations: Calculate (E20), via metrics algorithms distinct from the tracking algorithms T and stored in a memory of the collaborative tracking system, a metrics vector Target site of the target, the metric vector characterizing the target situation Target ID; Check if the metric vector Target site belongs to the reference area S ref T< of the tracking algorithm T; If yes, add the quintuplet ( Target ID , ID bestTracker , Target state , Target site , date) to a list to create (E21) the “ground truths” list Ground truth list, bestTracker ID denoting the tracking algorithm T of the triplet list included in the prediction dictionary Dict predictions and having the highest confidence score, in the calculation of the situation characterized by the vector of metrics Target site,compared to other scores obtained by other tracking algorithms of said prediction dictionary.

[0031] The “for each target” loop Target ID ....», above, means «for each target, establish as “ground truth”, for a tracking algorithm T, the state Target state most confident if T is in its reference zone S ref T< . The "ground truth" is therefore, in other words, the value taken by the state Target state predicted by the tracking algorithm T, included in the prediction dictionary Dict predictions and, having the highest confidence score, in calculating the situation characterized by the vector of metrics Target site , compared to other scores obtained by other tracking algorithms of said prediction dictionary.

[0032] In some embodiments, illustrated by the Figures 1A and 1Dand after implementation of the step of determining (E2) tracking algorithms T in the reference zone of each target resulting in the creation of the “ground truth” list List of ground truths, the step of updating (E3) chess areas is executed. Said updating step (E3) comprises

[0033] For each element ( Target ID, BestTracker ID , Target State GT< , Target site, date) from the “ground truth” list Ground truth list at least one of the steps consisting of: Search (E30), in the prediction dictionary for a given target Dict predictions, the prediction for the tracking algorithm T with the identifier ID tracker and the target Target ID on the given date. The prediction corresponds to the target state Target state in the list ( ID tracker, Target state , date) from the prediction dictionary Dict predictions for a given target, Calculate (E31) a likelihood score defined by the measurement of the distance between the target state and the “ground truth” list Ground truth list, GT target state<, and the target state searched in the list ( ID tracker, Target state , date) from the prediction dictionary Dict predictions for a given target, If the identity ID tracker of the T-tracking algorithm of the prediction dictionary Dict predictions, for a given target, is different from the identity ID bestTracker of the tracking algorithm T of the “ground truths” list Ground truth list (i.e. the tracking algorithm is not (and should not be) the "best tracking algorithm" that gave the state Target State GT< ) and the calculated likelihood score is less than a predetermined threshold value, then; Update (E32) the failure zone S fail T< of the tracking algorithm T by adding the metrics vector Target site calculated in the step of determining the tracking algorithm T in the reference area and included in the element element ( Target ID , ID bestTracker , Target State GT< , Target site, date) from the “ground truth” list Ground truth list Target State GT<corresponds to the target state of the best tracking algorithm T defined as “ground truth”. GT (Ground Truth) corresponds to the English meaning of the term “ground truth”.

[0034] For example, and without limitation, the distance between the two states Target State GT< And Target state , can be calculated by running a dedicated algorithm that compares the physical positions (in space) of the two states.

[0035] In certain embodiments, the dedicated algorithm may be, for example, and in a non-limiting manner, the Jaccard coefficient algorithm (or IoU method for Intersection over Union in English) which is a statistical method whose execution evaluates the similarity between the two states. Target State GT< And Target state. Similarity is defined by the Jaccard coefficient.

[0036] Thus, the failure zone of each tracking algorithm T of the single or multi-target tracking system is automatically updated according to the reference zones of the other tracking algorithms, i.e. in a collaborative manner. Updating the failure zone of a tracking algorithm, implemented by a tracking device (tracker) of the collaborative tracking system, i.e. the situations for which the tracking algorithm should not be trusted for the detection of a target, makes it possible to avoid the generation of erroneous information during the operation of said tracking device.

[0037] The invention also relates to a device for calibrating a single or multi-target tracking system.

[0038] In some embodiments, as illustrated in the Figure 2, the calibration device (1) can be integrated into the processing chain of a single or multi-target tracking or detection system. Said calibration device comprises at least one memory for storing a set of programs and at least one calculation unit for executing said programs to carry out the steps of the calibration method, as described in the present application, and storing the results of the calculations in the lists, ( List of ground truths, List of predictions dictionary Dict predictions containing each triplet ( ID tracker , Target state, date) of a target Target ID ), prediction dictionaries, reference area S ref T< or the chess area S fail T< known tracking algorithm T for each tracking device (or tracker) of a plurality of tracking devices, each tracking device comprising at least one tracking algorithm T.

[0039] There figure 3, illustrates the use of the calibration device in one embodiment. The calibration device (1) may be integrated into a collaborative tracking system comprising at least one set of cameras for capturing video images, a video stream server (VSS) for at least storing the video image streams captured by the cameras, and DT tracking devices implementing tracking algorithms that may be the same or different.The image stream from the video stream server (SFV) is sent or transmitted, on the one hand to an object detection device (RFCN), included in the tracking system, to deliver information for identifying new targets to each of the single-target tracking devices DT MIL and DT KCF executing MIL (Multiple Instance Learning) and KCF (kernelized correlation filter) type tracking algorithms, on the other hand to each of the tracking devices DT MIL and DT KCF whose output signals are sent for processing by the calibration device (1) as defined in the present application. The calibration device implements the calibration method, as defined in the present application, and comprises at least one device (DRCTS) executing a main collaborative tracking routine and dialoguing on the one hand with a situation calculator (CS) to obtain the metrics vector. Target sitecalculated for each new target and with a manager (GZE) of the known failure zones S fail T< to provide this information to a main tracking device (DTP) developing the tracking result from the information transmitted by the GZE manager and the information transmitted by the DT MIL or DT KCF tracking device.

[0040] In another embodiment illustrated by the Figure 4, The calibration device (1) can be integrated into a collaborative tracking system comprising at least one set of cameras for capturing video images, a video stream server (VSS) for at least storing the video image streams captured by the cameras and DT tracking devices implementing tracking algorithms which can be the same or different.The image stream from the video stream server is sent or transmitted, on the one hand to a multi-target tracking device DT MHT executing a tracking algorithm of the MHT (Multiple Hypothesis Tracking) type and an object detection device (RFCN) to deliver information for identifying or detecting new targets to said multi-target tracking device DT MHT, on the other hand to each of two sets of a plurality of tracking devices DT MIL and DT KCF executing tracking algorithms of the MIL (Multiple Instance Learning) and KCF (kernelized correlation filter) type, the multi-target tracking device D MHT transmitting target initialization information to each of the two sets of a plurality of tracking devices DT MIL and DT KCF. The information transmitted by the multi-target tracking device D MHT comprises at least the coordinates of the target(s).These coordinates can be used by the DT MIL and DT KCF tracking devices to extract, via algorithms stored in their memory, at least one thumbnail of the image around each target which will be used for their initialization. The output signals of each of the DT MHT, DT MIL, DT KCF tracking devices are sent for processing by the calibration device (1), as defined in the present application. Said calibration device (1) implements the calibration method, as described in the present application, and comprises at least one device (DRCTS) executing the main collaborative tracking routine (RCTS) and dialoguing on the one hand with a situation calculator (CS) to obtain the metrics vector. Target site calculated for each new target and with a manager (GZE) of the failure zones S fail T<known to provide this information to a main tracking device (DTP) developing the tracking result from the information transmitted by the GZE manager and the information transmitted by each DT MHT, DT MIL or DT KCF tracking device.

[0041] The tracking method as described in this application allows for reducing false positives of a global tracking system and thus improving global tracking.

[0042] By global tracking system we mean a tracking system using several tracking algorithms T, different or identical, to perform target tracking on the same video stream.

[0043] By defining a "tracklet" as a list of positions placed in time (x,y,t) that must refer to a target, for example, and without limitation, a person or the same physical object, we mean by false positive in a tracking process: the “tracklet” having within its list of positions positions not referring to any human or object of the ground truth The “tracklet” having positions belonging to different objects or humans.

[0044] For a "False Negative" situation (the "False Negative" situation corresponds to one where a target is present but not detected by a tracking algorithm T). In such a situation, the target can be detected by another tracking algorithm T' of the collaborative tracking system. This situation is not included in the failure zone S fail T< .

[0045] For a "False Positive" situation, the tracking algorithm T returns a response that does not match a given target. Two cases are possible: Either the situation is in the failure zone S fail T< and in this case, the response is not taken into account. Or the situation is not in the failure zone S fail T< of the tracking algorithm T and in this case the failure zone S fail T< of the tracking algorithm T is updated.

[0046] In another embodiment, the method of the present application can allow computational savings by making possible a strategy comprising the periodic switching off of a part of the tracking devices of a single or multi-target collaborative tracking system. For example, and in a non-limiting manner, a collaborative tracking system comprising at least two tracking devices DT A and DT B, each implementing respectively the tracking algorithms TA and TB and at least the calibration device (1) as defined in the present application. If we define by: S fail <none / > t i T , the failure zone comprising the known failure situations for a tracking algorithm T implemented by a tracking device DT at a time ti and; S FP <none / > t i T , an area containing the situations at a time ti, S fail <none / > t i T , excluded, for which the tracking algorithm T implemented by a tracking device DT gives bad positions for a target; said calibration device (1) implementing the calibration method, as defined in the present application, makes it possible to update the failure zone of the tracking algorithm T HAS at a time t 2 by calculating the component ( SREFt1TB∩SFPt1TA ) corresponding to the situations, at time t 1 , where the tracking algorithm T B is in the reference zone ( SREFt1TB ) and where the tracking algorithm T HAS is faulty and is not in the known failure zone, and adding it to the known failure zone Sfailt1TA at time t 1 : Sfailt2TA=Sfailt1TA+SREFt1TB∩SFPt1TA

[0047] Thus the system makes it possible to extend the known failure zone of the TA tracking algorithm between times t 1 and t 2 . The overall tracking system, comprising the tracking devices DT A and DT B , can then take advantage of the integration of the calibration device (1) implementing the calibration method when it is required to operate in two successive modes in which respectively: mode 1: tracking algorithms A and B work. We then have an update of S fail <none / > t 1 T A who becomes S fail <none / > t 2 T A ; mode 2: only tracking algorithm A is used. In this case, we use S fail <none / > t 2 T A

[0048] The tracking system then gives fewer false positives for all situations included in the area defined by S fail <none / > t 2 T A − S fail <none / > t 1 T A .

[0049] The second mode (mode 2) defined above can be achieved if: the TB tracking algorithm is not used due to the absence or temporary saving of computing resources the TB tracking algorithm is not used in a phase 2 after calibration (i.e. configuration) of TA in a phase 1. Phase 1 can be carried out in a different environment (laboratory, test bench) from the target environment or directly in the target environment.

[0050] The above example, provided for two tracking devices DT A and DT B each implementing the tracking algorithms TA and TB respectively, can be generalized directly for a global tracking system comprising N tracking devices DT Ai (i= 1, N), the update of the failure zones of the tracking algorithms being defined by: S fail <none / > t 2 T A 1 = S fail <none / > t 1 T A 1 + S REF <none / > t 1 T A 2 ∩ S FP <none / > t 1 T A 1 + … + S REF <none / > t 1 T A N ∩ S FP <none / > t 1 TA 1 ⋯ ⋯ ⋯ ⋯ S fail <none / > t 2 T A N = S fail <none / > t 1 T A N + S REF <none / > t 1 T A 1 ∩ S FP <none / > t 1 T A N + … + S REF <none / > t 1 T A N − 1 ∩ S FP <none / > t 1 TA N

[0051] The present application describes various technical features and advantages with reference to the figures and / or to various embodiments. Those skilled in the art will understand that the technical features of a given embodiment may in fact be combined with features of another embodiment unless the opposite is explicitly mentioned or it is obvious that these features are incompatible or that the combination does not provide a solution to at least one of the technical problems mentioned in the present application. Furthermore, the technical features described in a given embodiment may be isolated from the other features of this embodiment unless the opposite is explicitly mentioned.

[0052] It should be obvious to those skilled in the art that the present invention allows embodiments in many other specific forms without departing from the scope of the invention as claimed. Therefore, the present embodiments should be considered illustrative, but may be modified within the scope defined by the protection sought, and the invention should not be limited to the details given above.

Claims

1. Method for calibrating a single- or multi-target collaborative tracking system comprising at least one set of tracking algorithms T executed by at least one processor for processing the information from an image stream in order to detect targets in the video image stream which is captured at least by a given camera of the tracking system, according to a given viewpoint of a scene to be monitored, each tracking algorithm T using at least one known reference zone SrefT representing the set of situations for which the algorithm detects an actual target and at least one known failure zone SfailT representing the set of situations for which the tracking algorithm does not detect an actual target, the single- or multi-target tracking system comprising at least one set of modules / programs stored in a memory and at least one processor executing the set of modules to implement the calibration method, said method being characterized in that it comprises at least the steps of: - collecting (E1) the combinations of all the tracking algorithms T of the single- or multi-target tracking system in order to create a dictionary of predictions Dictpredictions for a given target; - determining (E2), by means of the created predictions dictionary Dictpredictions, the tracking algorithms T in the reference zone in order to create a "ground truths" list ListGround truths comprising the tracking algorithms T which have a confidence score greater than a predefined threshold value; - updating (E3) the known failure zone of each tracking algorithm T of the list of tracking algorithms T of the single- or multi-target tracking system not belonging to the "ground truths" list ListGround truths and having a likelihood score below a predetermined threshold.

2. Method for calibrating a single- or multi-target collaborative tracking system according to claim 1, characterized in that the step of collecting tracking algorithm combinations consists of at least: For each tracking algorithm T: - predicting (E10), from the video image stream, the states of a target or a set of targets contained in the image streams in order to create a predictions dictionary DictpredictionsT linked to the tracking algorithm T, said predictions dictionary DictpredictionsT taking as input an existing or new target identifier IDtarget and giving as output as a value a vector characterizing the state of the new target, called Statetarget; - for each triplet (IDtarget, Statetarget, date) included in the predictions dictionary DictpredictionsT of the tracking algorithm T, creating (E11) the predictions dictionary Dictpredictions for a given target by adding the triplet (IDtracker, Statetarget, date) to a list, the prediction dictionary Dictpredictions for a given target taking as input the target identifier IDtarget and giving as the output value a triplet list (IDtracker, Statetarget, date) where IDtracker is the identifier of the tracking algorithm T and date is the value of the date on which the target was detected by the tracking algorithm T.

3. Method for calibrating a single- or multi-target tracking system according to claims 1 and 2, characterized in that the step (E2) of determining tracking algorithms T in a reference zone comprises: For each target IDtarget and triplet list (IDtracker, Statetarget, date) in the predictions dictionary Dictpredictions of a target: - calculating (E20) a metrics vector Sittarget of the target, the metric vector characterizing the situation of the target IDtarget; - checking whether the metric vector Sittarget belongs to the reference zone SrefT of the tracking algorithm T; - if yes, adding the quintuplet (IDtarget, IDbestTracker, Statetarget, Sittarget, date) to a list in order to create (E21) the "ground truths" list ListGround truths, IDbestTracker designating the tracking algorithm T of the triplet list included in the predictions dictionary Dictpredictions and having the highest confidence score, in calculating the situation characterized by the metrics vector Sittarget, in comparison with the other scores obtained by the other algorithms of said predictions dictionary.

4. Method for calibrating a single- or multi-target tracking system according to claims 1 and 2, characterized in that the updating step (E3) comprises: For each element (IDtarget, IDbestTracker, StatetargetGT, Sittarget, date) of the "ground truths" list List Ground truths at least one of the steps consisting of: - searching (E30), in the predictions dictionary Dictpredictions for a given target, the prediction for the tracking algorithm T with the identifier IDtracker and target IDtarget on the given date, the prediction corresponds to the target state Statetarget in the list (IDtracker, Statetarget, date) of the predictions dictionary Dictpredictions; - calculating (E31) a likelihood score defined by measuring the distance between the target state of the "ground truths" list List Ground truths, StatetargetGT, and the target state searched in the list (IDtracker, Statetarget, date) of the predictions dictionary Dictpredictions; - if the identity IDtracker of the tracking algorithm T of the predictions dictionary Dictpredictions is different from the identity IDbestTracker of the tracking algorithm T of the "ground truths" list ListGround truths and the calculated likelihood score is less than a predetermined threshold value, then; - updating (E32) the failure zone SfailT of the tracking algorithm T by adding thereto the metrics vector Sittarget calculated in the step of determining the tracking algorithm T in the reference zone and included in the element element (IDtarget, IDbestTracker, StatetargetGT, Sittarget, date) of the"ground truths" list ListGround truths.

5. Calibration device (1) for a single- or multi-target tracking system, characterized in that it comprises at least one memory for storing a set of programs and at least one computing unit for executing said programs to carry out the steps of the calibration method according to any one of the preceding claims and to store the results of the computing calculations in the lists, (ListGround truths, Predictions dictionary list Dictpredictions containing each triplet (IDtracker, Statetarget, date) of a target IDtarget), the predictions dictionaries, the reference zone SrefT or the known failure zone SfailT of the tracking algorithm T for each tracking device (tracker) of a plurality of tracking devices of the tracking system, each tracking device comprising at least one tracking algorithm T.

6. Use of the calibration method of a single- or multi-target collaborative tracking system according to one of claims 1 to 4, characterized in that the image stream of a video stream server (SFV) is sent both to an object detection device (RFCN) for delivering new target identification information to each of the single-target tracking devices DTMIL and DTKCF executing MIL (Multiple Instance Learning) and KCF (kernelized correlation filter) type tracking algorithms, and to each of the tracking devices DTMIL and DTKCF, the output signals of which are sent for processing by the calibration device (1), according to claim 5, implementing the calibration method, according to claims 1 to 4, and comprising a device (DRCTS) executing a main collaborative tracking routine and communicating with a situations computer (CS) in order to obtain the metrics vector Sittarget calculated for each new target and with a manager (GZE) of the known failure zones SfailT in order to supply this information to a main tracking device (DTP) producing the tracking result from the information transmitted by GZE and the information transmitted by the tracking device DTMIL or DTKCF.

7. Use of the calibration method of a single- or multi-target collaborative tracking system according to one of claims 1 to 4, characterized in that the image stream of a video stream server is sent both to a multi-target tracking device DTMHT executing an MHT (Multiple Hypothesis Tracking) type tracking algorithm and an object detection device (RFCN) for delivering new target identification information to said multi-target tracking device DTMHT, and to each of two sets of a plurality of tracking devices DTMIL and DTKCF executing MIL (Multiple Instance Learning) and KCF (kernelized correlation filter) type tracking algorithms, the multi-target tracking device DMHT transmitting target initialization information comprising at least the coordinates of said targets to each of the two sets of a plurality of tracking devices DTMIL and DTKCF, the output signals of each of the tracking devices DTMHT, DTMIL, DTKCF are sent for processing by the calibration device (1), according to claim 5, implementing the calibration method, according to claims 1 to 4, and comprising a device (DRCTS) executing the main collaborative tracking routine (RCTS) and communicating with a situations (Sit) computer (CS) in order to obtain the metrics vector Sittarget calculated for each new target and with a manager (GZE) of known the failure zones SfailT to supply this information to a main tracking device (DTP) producing the tracking result from the information transmitted by GZE and the information transmitted by each tracking device DTMHT, DTMIL or DTKCF.