A cross-temporal and spatial object association method based on the fusion of visual features and spatiotemporal information

Abstract

The present invention relates to the field of image recognition technology, and in particular to a cross-time and space target association method that fuses visual features with time and space information, comprising the following steps: extracting target features of surveillance videos and images to be associated; constructing a dynamic time and space topological structure between cameras in a large time and space; screening the time and space association range based on the dynamic time and space topological structure between cameras; using target features and the principle of time and space consistency to determine whether the target image to be associated and the image within the association range have the same target identity, and then performing target association in a small time and space; using the associated target image as a new image to be associated, gradually expanding the search range, and repeatedly performing the above small time and space range association to achieve target association in a large time and space range. The present invention can fuse target features and the time and space information between cameras to achieve more efficient and accurate large time and space target association.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and more specifically to a cross-temporal and spatial target association method that fuses visual features with temporal and spatial information. Background Art

[0002] In the content analysis of surveillance videos, associating the time and location clues of the same target to form the direction of the pedestrian's progress has become a basic task of intelligent surveillance systems. A popular approach is to use content-based image retrieval technology, use the image of the target person as the query image, perform image retrieval in the camera network, find other images of the person, and determine the time and place of the person's appearance with the help of the camera's geographical location and the video shooting time. This is pedestrian re-identification technology.

[0003] At present, pedestrian re-identification has become one of the important basic technologies for intelligent analysis of surveillance videos. Person re-identification mainly uses the apparent features of the human body for similarity matching, such as the color, style, and texture of clothing. In a large spatiotemporal scale, there are a large number of different pedestrians with similar appearances that are incorrectly associated, and the same pedestrian cannot be associated due to different viewing angles, different directions, different lighting, etc.

[0004] In view of the problem that the visual appearances of different targets are similar and the visual appearances of the same target vary greatly at large temporal and spatial scales, how to comprehensively consider the target characteristics and temporal and spatial position information to achieve target association over a large temporal and spatial range is a technical problem that needs to be urgently solved in this field. Summary of the invention

[0005] In view of this, the present invention provides a cross-temporal and spatial target association method by fusing visual features with spatiotemporal information, which can fuse target features and the spatiotemporal information between cameras to achieve more efficient and accurate large-scale spatiotemporal target association.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A cross-temporal and spatial target association method by fusing visual features with spatiotemporal information comprises the following steps:

[0008] Step 1: Detect and track targets in surveillance videos based on pre-trained target recognition and tracking models, and extract target features;

[0009] Step 2: Count the distance and transit time of the target between different cameras at different time periods, as well as the situation of the target leaving the camera's field of view under different cameras, and construct a dynamic spatiotemporal topological structure between cameras in a large spatiotemporal range;

[0010] Step 3: For a given target image or image sequence to be associated, the spatiotemporal association range of the target image or image sequence to be associated across cameras is screened by utilizing the dynamic spatiotemporal topological structure between the cameras;

[0011] Step 4: Within the spatiotemporal association range determined in step 3, target features and spatiotemporal consistency are used as criteria to perform target association on the given target image or image array to be associated in a small spatiotemporal range; if the target cannot be associated, return to step 3 to redetermine the spatiotemporal association range; if the target is associated, execute step 5;

[0012] Step 5: Using the target image or image sequence associated in step 4 as the image to be associated, repeat steps 3 to 5, gradually expanding the target search range until target association is completed in a large spatiotemporal range.

[0013] Furthermore, the target in step 1 is a pedestrian or a vehicle. When it is a pedestrian, the extracted target features include facial features and body features that can distinguish the identity of the pedestrian, wherein the body features include: head features, visual appearance features, attribute features or body contour features.

[0014] Furthermore, step 2 includes:

[0015] Step 21: For n cameras in the large spatiotemporal scene, construct a dynamic spatiotemporal topological structure between the n cameras and represent it with an n-order topological relationship matrix A;

[0016] Each element in the matrix A consists of a pair, expressed as: A(i, j, k)=(t k , d ij , t ij ), where t k The current time period is divided into weekday morning and evening peaks, weekday non-peaks, weekends, and other custom special periods; d ij The statistically obtained t k The physical distance of the target moving from the i-th camera to the j-th camera in the time period, t ij The statistically obtained t k The shortest transit time of the target from the i-th camera to the j-th camera in the time period;

[0017] Step 22: Use the target recognition and tracking model to monitor in real time whether a target leaves or enters the camera's field of view under each camera, and build a target departure list L for each camera in real time. If it is detected that a target leaves the field of view of the i-th camera at time t, set L[i, t, 1] to 1, otherwise set it to 0.

[0018] Further, in step 21, t is obtained by using the API interface provided by Amap, Baidu or other open map / GIS platforms. k The physical distance and shortest flight time of the target moving from the i-th camera to the j-th camera in the time period.

[0019] Furthermore, step 3 includes:

[0020] Step 31: for a given target image or image sequence to be associated and its time and space position information, according to the dynamic spatiotemporal topological structure constructed in step 21, screen the camera where the target is located and the cameras whose physical distance from the camera is less than the initial preset distance threshold D, and use the screened camera ranges as the preliminary association range R0;

[0021] Step 32: Based on the target departure list L of each camera constructed in real time in step 22, filter out the camera range R1 that perceives the target out of the field of view as of the current moment, where R1 = the camera range of the initial association range R0 – {L[i, t, 1]=0};

[0022] Step 33: Based on the camera range R1, select from all the target images or image sequences detected and tracked that the time interval with the target image or image sequence to be associated is greater than or equal to the shortest transit time t between the corresponding cameras. ij , but less than k t t ij Pedestrian images or image sequences are used as the spatiotemporal correlation range; where k t is the time domain expansion coefficient, and k t >1.

[0023] Furthermore, step 3 also includes:

[0024] Step 34: When the spatiotemporal correlation range determined in step 33 cannot be associated with the target, the temporal search range in step 33 is expanded to the original k t times (when the maximum time search range in the time domain exceeds a certain threshold, such as 10 times the shortest transit time between the original corresponding cameras t ij When the time domain expansion stops, the expanded space-time correlation range is re-determined;

[0025] Step 35: When the spatiotemporal correlation range determined in step 34 cannot be associated with the target, the initial preset distance threshold D in step 31 is expanded to k s D, k s is the spatial expansion coefficient, and k s >1; repeat steps 31 to 34 to redefine the expanded spatiotemporal correlation range.

[0026] Furthermore, in step 31, for pedestrian targets, the value range of the initial preset distance threshold D is 500 meters to 1000 meters; for vehicle targets, the value range of the initial preset distance threshold D is 1500 meters to 3000 meters.

[0027] Furthermore, when the target is a pedestrian, step 4 includes:

[0028] Step 41: extracting pedestrian facial features and pedestrian visual appearance features in a given pedestrian image or image sequence to be associated, as well as pedestrian facial features and pedestrian visual appearance features within the spatiotemporal association range in step 3 based on the target recognition and tracking model;

[0029] Step 42: compare the facial features of the pedestrians in the pedestrian image or image sequence to be associated and the facial features of the pedestrians in the spatiotemporal association range determined in step 3 with the faces in the database. If both can be identified based on the face and the identities are the same, they are determined to be the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or sequence in the spatiotemporal association range; if the identities are different, they are determined to be different pedestrians and no association is performed.

[0030] Step 43: If the identity of the pedestrian image or image sequence to be associated or within the spatiotemporal association range cannot be identified in step 42, the facial feature similarity between the pedestrian image or image sequence to be associated and the pedestrian image or image sequence within the spatiotemporal association range determined in step 3 is calculated. If the similarity is higher than a preset threshold, it is determined that they are the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or image sequence within the spatiotemporal association range; otherwise, it is determined that they are different pedestrians and no association is performed;

[0031] Step 44: When facial features are not available, calculate the similarity of the pedestrian visual appearance features between the pedestrian image or image sequence to be associated and the pedestrian image or sequence within the spatiotemporal association range determined in step 3. If the similarity is higher than a preset threshold, they are determined to be the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or sequence within the spatiotemporal association range; otherwise, they are determined to be different pedestrians and no association is performed.

[0032] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses a cross-time and space target association method that fuses visual features and time and space information. First, the target features of the monitoring video and the image to be associated are extracted; then, the dynamic time and space topological structure between the cameras in a large time and space is constructed; secondly, the time and space association range is screened based on the dynamic time and space topological structure between the cameras; thirdly, the target features and the principle of time and space consistency are used to determine whether the target image to be associated and the image in the association range have the same target identity, and then the target association is performed in a small time and space; finally, the associated target image is used as the new image to be associated, the search range is gradually expanded, and the above association is repeated to achieve target association in a large time and space range. The present invention can achieve more efficient and accurate large time and space target association by fusing target features and the time and space information between each camera. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0034] Figure 1 A flow chart of a cross-temporal and spatial target association method that fuses visual features and spatiotemporal information provided by the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] like Figure 1 As shown, the embodiment of the present invention discloses a method for associating cross-temporal and spatial targets by fusing visual features with spatiotemporal information, comprising the following steps:

[0037] Step 1: Detect and track targets in surveillance videos based on pre-trained target recognition and tracking models, and extract target features;

[0038] Step 2: Count the distance and transit time of the target between different cameras at different time periods, as well as the situation of the target leaving the camera's field of view under different cameras, and construct a dynamic spatiotemporal topological structure between cameras in a large spatiotemporal range;

[0039] Step 3: For a given target image or image sequence to be associated, the spatiotemporal association range of the target image or image sequence to be associated across cameras is screened by utilizing the dynamic spatiotemporal topological structure between the cameras;

[0040] Step 4: Within the spatiotemporal association range determined in step 3, target features and spatiotemporal consistency are used as criteria to perform target association on the given target image or image array to be associated in a small spatiotemporal range; if the target cannot be associated, return to step 3 to redetermine the spatiotemporal association range; if the target is associated, execute step 5;

[0041] Step 5: Use the target image or image sequence associated in step 4 as a new image to be associated, repeat steps 3 to 5, and gradually expand the target search range until the target association is completed in a large temporal and spatial range.

[0042] In a specific embodiment, the target in step 1 is a pedestrian or a vehicle. When it is a pedestrian, the extracted target features include facial features and body features that can distinguish the identity of the pedestrian, wherein the body features include: traditional pedestrian re-identification features, head features, visual appearance features, attribute features or body contour features.

[0043] Specifically, the target recognition and tracking model in step 1 is composed of multiple deep learning networks. Taking pedestrian detection as an example, the target recognition and tracking model includes: target detection and tracking network model, face detection and recognition network model and human feature extraction network model. Specifically, various deep learning models such as ResNet and ViT can be selected.

[0044] For the selected network model, based on the annotation information of the training sample images and the back propagation algorithm, supervise and complete the training of the above network model;

[0045] Finally, based on the trained network model, pedestrian detection and tracking are achieved in the surveillance video, and facial features and visual appearance features of pedestrians that can help distinguish the identity of pedestrians are extracted from pedestrian images.

[0046] In one embodiment, step 2 includes:

[0047] Step 21: For n cameras in the large spatiotemporal scene, construct a dynamic spatiotemporal topological structure between the n cameras and represent it with an n-order topological relationship matrix A;

[0048] Each element in the matrix A consists of a pair, expressed as: A(i, j, k)=(t k , d ij , t ij ), where t k The current time period is divided into weekday morning and evening peaks, weekday non-peaks, weekends, and other custom special periods; dij The statistically obtained t k The physical distance of the target moving from the i-th camera to the j-th camera in the time period, t ij The statistically obtained t k The shortest transit time of the target from the i-th camera to the j-th camera in the time period; obtain t by using the API interface provided by Amap, Baidu or other open map / GIS platforms k The physical distance and shortest flight time of the target moving from the i-th camera to the j-th camera in the time period.

[0049] Step 22: Use the target recognition and tracking model to monitor in real time whether a target leaves or enters the camera's field of view under each camera, and build a target departure list L for each camera in real time. If it is detected that a target leaves the field of view of the i-th camera at time t, set L[i, t, 1] to 1, otherwise set it to 0.

[0050] In a specific embodiment, step 3 includes:

[0051] Step 31: for a given target image or image sequence to be associated and its time and space position information, according to the dynamic spatiotemporal topological structure constructed in step 21, screen the camera where the target is located and the cameras whose physical distance from the camera is less than the initial preset distance threshold D, and use the screened camera ranges as the preliminary association range R0;

[0052] For pedestrian targets, the initial preset distance threshold D ranges from 500 meters to 1000 meters; for vehicle targets, the initial preset distance threshold D ranges from 1500 meters to 3000 meters.

[0053] Step 32: Based on the target departure list L of each camera constructed in real time in step 22, filter out the camera range R1 that perceives the target out of the field of view as of the current moment, where R1 = the camera range of the initial association range R0 – {L[i, t, 1]=0};

[0054] Step 33: Based on the camera range R1, select from all the target images or image sequences detected and tracked that the time interval with the target image or image sequence to be associated is greater than or equal to the shortest transit time t between the corresponding cameras. ij , but less than k t t ij Pedestrian images or image sequences are used as the spatiotemporal correlation range; where k t is the time domain expansion coefficient, and k t >1.

[0055] Steps 31 and 32 are to determine the airspace search range, that is, the camera range, under which cameras to search; step 33 is to confirm within which time period to search under these cameras.

[0056] Step 34: When the spatiotemporal correlation range determined in step 33 cannot be associated with the target, the temporal search range in step 33 is expanded to the original k t times (when the maximum time search range in the time domain exceeds a certain threshold, such as 10 times the shortest transit time between the original corresponding cameras t ij , stop time domain expansion), and redefine the expanded space-time correlation range.

[0057] Step 35: When the spatiotemporal correlation range determined in step 34 cannot be associated with the target, the initial preset distance threshold D in step 31 is expanded to k s D, k s is the spatial expansion coefficient, and k s >1; repeat steps 31-34 to redefine the expanded spatiotemporal correlation range.

[0058] Time domain expansion coefficient k t and the time domain expansion coefficient k s The value of k can generally be 1.5 or 2 according to experience. t and k s The larger it is, the faster the search range expands, the larger the corresponding target image data to be associated is, and the larger the amount of feature similarity calculation is.

[0059] In a specific embodiment, when the target is a pedestrian, step 4 includes:

[0060] Step 41: extracting pedestrian facial features and pedestrian visual appearance features in a given pedestrian image or image sequence to be associated, as well as pedestrian facial features and pedestrian visual appearance features within the spatiotemporal association range in step 3 based on the target recognition and tracking model;

[0061] Step 42: The facial features of the pedestrians in the pedestrian image or image sequence to be associated and the facial features of the pedestrians within the spatiotemporal association range determined in step 3 are compared with the faces in the database respectively. If both can be identified based on the face and the identities are the same, they are determined to be the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or sequence within the spatiotemporal association range; if the identities are different, they are determined to be different pedestrians and no association is performed; in this step, face recognition refers to directly determining the identity of the pedestrian with the face through 1:N face comparison.

[0062] Step 43: If the identity of the pedestrian image or image sequence to be associated or within the spatiotemporal association range cannot be identified in step 42, the facial feature similarity between the pedestrian image or image sequence to be associated and the pedestrian image or image sequence within the spatiotemporal association range determined in step 3 is calculated. If the similarity is higher than a preset threshold, it is determined that they are the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or image sequence within the spatiotemporal association range; otherwise, it is determined that they are different pedestrians and no association is performed;

[0063] In this step, the facial feature similarity comparison refers to 1:1 face verification, that is, verifying whether the face of the pedestrian image to be associated is the same as the face in each image within the associated range.

[0064] Step 44: When facial features are not available, calculate the similarity of the pedestrian visual appearance features between the pedestrian image or image sequence to be associated and the pedestrian image or sequence within the spatiotemporal association range determined in step 3. If the similarity is higher than a preset threshold, they are determined to be the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or sequence within the spatiotemporal association range; otherwise, they are determined to be different pedestrians and no association is performed.

[0065] In this step, facial recognition is limited to reasonable scenarios, such as those where it can be presumed based on common sense that the person being recognized is aware of the situation, has authorized it, and can choose to use it, such as attendance at work, entry and exit of parks, and ticket checking at stations and scenic spots. At the same time, facial recognition in this application is limited to legal scenarios: such as finding lost children.

[0066] In this step, the pedestrian visual appearance feature similarity can be calculated using discrete cosine distance. The discrete cosine distance calculation formula of feature A and feature B is as follows:

[0067] ;

[0068] Among them, n is the feature dimension, A i and B i Represents the features of the i-th dimension of the feature vector.

[0069] In addition, the principles of the present invention can also be used to analyze video image targets such as vehicles, ships, and aircraft acquired at different times and spaces and from different viewing angles, and to achieve large-scale spatiotemporal target association by fusing target visual features with spatiotemporal reasoning.

[0070] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0071] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cross-temporal and spatial target association method by fusing visual features with spatiotemporal information, characterized in that: The following steps are involved: Step 1: Detect and track targets in surveillance videos based on pre-trained target recognition and tracking models, and extract target features; Step 2: Count the distance and transit time of the target between different cameras at different time periods, as well as the situation of the target leaving the camera's field of view under different cameras, and construct a dynamic spatiotemporal topological structure between cameras in a large spatiotemporal range; Step 3: For a given target image or image sequence to be associated, the spatiotemporal association range of the target image or image sequence to be associated across cameras is screened by utilizing the dynamic spatiotemporal topological structure between the cameras; Step 4: within the spatiotemporal association range determined in step 3, target features and spatiotemporal consistency are used as criteria to perform target association on the given target image or image array to be associated in a small spatiotemporal range; If the target cannot be associated, return to step 3 to re-determine the time and space association range. If the target can be associated, execute step 5; Step 5: Using the target image or image sequence associated in step 4 as the image to be associated, repeat steps 3 to 5, gradually expanding the target search range until target association is completed in a large temporal and spatial range; Wherein, step 2 comprises: Step 21: Large space-time scene n Cameras, build n The dynamic spatiotemporal topology of the cameras is n Order topological relationship matrix A express; matrix A Each element in is composed of a two-tuple, represented as: A ( i , j , k )=( t k , d ij , t ij ),in, t k The current time period is divided into weekday morning and evening peaks, weekday non-peaks, weekends, and other custom special periods; d ij Indicates the statistical t k In the time period, the target is from i The camera moves to j The physical distance of the camera, t ij Indicates the statistical t k In the time period, the target is from i The camera moves to j The shortest transit time for each camera; Step 22: Use the target recognition and tracking model to monitor in real time whether a target leaves or enters the camera's field of view under each camera, and build a target departure list for each camera in real time L If the first i Cameras in t If the target leaves the field of vision, L [ i , t , 1] is set to 1, otherwise it is set to 0.

2. The method for associating objects across time and space by fusing visual features with time and space information according to claim 1, characterized in that: The target in step 1 is a pedestrian or a vehicle. When it is a pedestrian, the extracted target features include facial features and body features that can distinguish the identity of the pedestrian, where the body features include: head features, visual appearance features, attribute features or body contour features.

3. The method for associating objects across time and space by fusing visual features with time and space information according to claim 1, characterized in that: In step 21, the API interface provided by the open map / GIS platform is used to obtain t k In the time period, the target is from i The camera moves to j The physical distance and shortest flight time of the cameras.

4. The method for associating objects across time and space by fusing visual features with time and space information according to claim 1, characterized in that: Step 3 includes: Step 31: For a given target image or image sequence to be associated and its temporal and spatial location information, based on the dynamic spatiotemporal topological structure constructed in step 21, filter the camera where the target is located and the camera whose physical distance from the camera is less than the initial preset distance threshold. D The selected camera ranges are used as the initial association ranges. R 0; Step 32: Build a list of target departures for each camera in real time based on step 22 L , filter out the camera range that perceives the target as out of view at the current moment R 1, where R 1 = Preliminary association range R 0 – { L [ i , t , 1]=0} camera range; Step 33: In Camera Range R 1, from all the target images or image sequences obtained by detection and tracking, select the target images or image sequences whose time interval with the target image or image sequence to be associated is greater than or equal to the shortest transit time between the corresponding cameras. t ij , but less than k t t ij Pedestrian images or image sequences are used as the spatiotemporal correlation range; k t is the time domain expansion coefficient, and k t >1.

5. The method for associating objects across time and space by fusing visual features with time and space information according to claim 4, characterized in that: Step 3 also includes: Step 34: When the time-space correlation range determined in step 33 cannot be associated with the target, the time domain search range in step 33 is expanded to the original k t times, redefine the expanded scope of space-time correlation; Step 35: When the spatiotemporal correlation range determined in step 34 cannot be associated with the target, the initial preset distance threshold in step 31 is D Expand to k s D , k s is the spatial expansion coefficient, and k s >1; repeat steps 31 to 34 to redefine the expanded spatiotemporal correlation range.

6. The method for associating objects across time and space by fusing visual features with time and space information according to claim 4, characterized in that: In step 31, for pedestrian targets, the initial preset distance threshold D The value range of is 500 meters to 1000 meters; for vehicle targets, the value range of the initial preset distance threshold D is 1500 meters to 3000 meters.

7. The method for associating objects across time and space by fusing visual features with time and space information according to claim 1, characterized in that: When the target is a pedestrian, step 4 includes: Step 41: extracting pedestrian facial features and pedestrian visual appearance features in a given pedestrian image or image sequence to be associated, as well as pedestrian facial features and pedestrian visual appearance features within the spatiotemporal association range in step 3 based on the target recognition and tracking model; Step 42: compare the facial features of the pedestrians in the pedestrian image or image sequence to be associated and the facial features of the pedestrians in the spatiotemporal association range determined in step 3 with the faces in the database. If both can be identified based on the face and the identities are the same, they are determined to be the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or sequence in the spatiotemporal association range; if the identities are different, they are determined to be different pedestrians and no association is performed. Step 43: If the identity of the pedestrian image or image sequence to be associated or within the spatiotemporal association range cannot be identified in step 42, the facial feature similarity between the pedestrian image or image sequence to be associated and the pedestrian image or image sequence within the spatiotemporal association range determined in step 3 is calculated. If the similarity is higher than a preset threshold, it is determined that they are the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or image sequence within the spatiotemporal association range; otherwise, it is determined that they are different pedestrians and no association is performed; Step 44: When facial features are not available, calculate the similarity of the pedestrian visual appearance features between the pedestrian image or image sequence to be associated and the pedestrian image or sequence within the spatiotemporal association range determined in step 3. If the similarity is higher than a preset threshold, they are determined to be the same pedestrian, and the pedestrian image or image sequence to be associated is deemed to be successfully associated with the pedestrian image or sequence within the spatiotemporal association range; otherwise, they are determined to be different pedestrians and no association is performed.

Images