A multi-source monitoring image integrated abnormal behavior personnel identification method

Abstract

The present application belongs to the technical field of abnormal behavior personnel identification, and particularly relates to a multi-source monitoring image integrated abnormal behavior personnel identification method. The steps include: acquiring original video data of multiple cameras in a target area; judging complete appearance and partial disappearance of personnel through a posture estimation and semantic segmentation model, taking the same as the start and end to perform inter-frame sampling and retain space-time information; pre-processing the image frames, extracting static posture and dynamic behavior features; fusing dynamic behavior, appearance and position information, realizing cross-camera personnel identity association and constructing behavior trajectory sequences; matching the trajectories with preset templates, combining time consistency, position path similarity, action matching degree and abnormal segment confidence aggregation value, and calculating deviation score to identify abnormal personnel. The method breaks through the limitation of single perspective, improves low-illumination recognition accuracy, solves the cross-perspective identity fracture problem, and enhances the accuracy and traceability of abnormal behavior identification.

Description

Technical Field

[0001] This invention belongs to the field of abnormal behavior identification technology, and in particular relates to a method for abnormal behavior identification by integrating multi-source surveillance images. Background Technology

[0002] With the accelerated construction of smart cities, public safety, and security prevention and control systems, intelligent behavior recognition technology based on video surveillance has become one of the key technical means in the practical application of public security video images. Existing technologies typically employ video image recognition models based on a single viewpoint to identify the behavior of individuals appearing in the surveillance footage and issue anomaly warnings. However, in real-world scenarios, issues such as varying monitoring point deployment angles, complex personnel behavior paths, severe occlusion, and difficulties in cross-viewpoint identity association arise, making the identification results of abnormal behavior susceptible to interference and posing risks of missed or false judgments. Furthermore, some behaviors exhibit phased and chain-like characteristics, making it difficult to reconstruct behavioral intent based solely on a single frame or viewpoint. This is especially true in areas covered by multiple cameras, where inconsistencies in the identification results of the same person under different cameras further reduce the system's accuracy in assessing behavioral risks. On the other hand, traditional abnormal behavior recognition methods largely rely on the output results of action classification models, lacking semantic-level linkage analysis with behavioral trajectories, regional rules, and behavioral deviation patterns, making it difficult to support the needs of public security in tracing personnel behavior trajectories, reconstructing behavioral chains, and identifying intent in practical applications. Summary of the Invention

[0003] To address the technical problems existing in the background art described above, this invention proposes a method for identifying abnormal behavior personnel by integrating multi-source surveillance images.

[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0005] S1. Video Data Acquisition: Acquire raw video data from multiple surveillance cameras deployed within the target monitoring area. The video data includes image streams from different perspectives and time periods.

[0006] S2. Video image frame extraction: Perform inter-frame sampling on the original video stream, extract the image frame sequence, and retain the timestamp, camera number, and location information corresponding to each frame image;

[0007] S3. Image preprocessing and feature extraction: Perform preprocessing operations on the image frame sequence and extract the behavioral features of human targets in the image frame sequence;

[0008] S4. Personnel Trajectory Reconstruction and Identity Association: Spatiotemporal identity matching will be performed on the same person identified under different cameras to construct a cross-view personnel behavior trajectory sequence;

[0009] S5. Behavioral rule deviation analysis: Match the personnel's behavioral trajectory with the preset template, calculate the deviation score of abnormal behavior, and if the deviation score of abnormal behavior exceeds the set deviation score threshold, the personnel are identified as abnormal behavior personnel.

[0010] Preferably, in step S2, the operation of inter-frame sampling of the original video stream is to first determine that the person has completely appeared in the video stream as the start of inter-frame sampling, and determine that the person has partially disappeared as the end of inter-frame sampling.

[0011] Preferably, the method for determining whether something is completely present or partially absent is as follows:

[0012] Key point network acquisition: Employs a pose estimation model to acquire key human body nodes in real time;

[0013] Preliminary complete appearance judgment: When all the collected key nodes are located within the effective area of ​​the image, and the confidence of the key points is higher than the set confidence threshold, it is judged as preliminary complete appearance;

[0014] Preliminary partial disappearance judgment: Define the importance weight of key nodes and calculate the weighted ratio of disappearing nodes. When the disappearance ratio exceeds the set threshold, it is judged as preliminary partial disappearance.

[0015] Enhanced judgment effect: A human semantic segmentation model is used to segment the pixel region of the human body and calculate the coverage rate of the human body mask and the image region. When the coverage rate is ≥90%, the auxiliary confirmation will judge the initial complete appearance as complete appearance; when the coverage rate is ≤40%, the auxiliary confirmation will judge the initial partial disappearance as partial disappearance.

[0016] Preferably, step S3, image preprocessing and feature extraction, includes:

[0017] Image enhancement processing is performed on the image frame, including image contrast enhancement, illumination normalization processing and edge denoising processing, to improve the recognition accuracy of human body regions in low-light environments.

[0018] A static pose feature vector is constructed using the key human node information of each frame image.

[0019] Furthermore, by combining the changing trends of key points between adjacent image frames, dynamic behavioral features of the human body are extracted to describe the human body's action state in the image frame.

[0020] Preferably, step S4, personnel trajectory reconstruction and identity association, includes:

[0021] By utilizing human dynamic behavior characteristics and fusing appearance features with location information, a joint feature representation for cross-camera identity matching is formed.

[0022] In multiple camera image frame sequences, the identities of people appearing under different cameras are associated based on joint feature representation to determine whether they belong to the same target individual;

[0023] Based on successful identity verification, a sequence of personnel behavior trajectories is constructed according to the time sequence, spatial location, and actions of personnel appearing under different cameras.

[0024] Preferably, step S5, which matches the personnel's behavioral trajectory with a preset template and calculates the deviation score of abnormal behavior, includes the following specific operations:

[0025] The behavioral trajectory of a person is encoded into a time-ordered behavioral sequence, which consists of location nodes, behavioral actions, and behavioral times.

[0026] Select the normal behavior template sequence that is closest to the person's behavior type and path structure, and calculate the trajectory matching score S based on temporal consistency, location path similarity, and action sequence matching degree. match The calculation method for time consistency is as follows: Where N is the trajectory length, T t T represents the timestamp at time t. t norm The timestamp represents the template time t; the method for calculating location path similarity is as follows: Where M represents the total number of frames in the normal template, L t This represents the positional characteristics at time t. The positional features at time t of the template are represented; the calculation method for the degree of action sequence matching is as follows: Where K represents the number of critical action points, and I() is an indicator function that returns 1 if the condition is met and 0 otherwise. The behavioral characteristics at time t are represented by these characteristics. The template time t represents the behavioral and action features; the trajectory matching score is obtained by weighted fusion of the three.

[0027] Multiple potential anomalous behavior segments are identified in the behavioral trajectory, and each anomalous behavior segment is assigned an anomalous behavior confidence level C. j The aggregate confidence value S for anomalous behavior is calculated by combining the duration. abn The confidence aggregation value S abn The calculation method is as follows: Where J represents the number of potential anomalous behavior segments identified;

[0028] Finally, the trajectory matching score and the anomaly aggregation value are combined to construct the final deviation score S. deviation The calculation method is as follows: S deviation =(1-S match )·Sabn .

[0029] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0030] 1. Employing multi-camera collaborative acquisition of video data from different perspectives and time periods overcomes the limitations of single-view information. By combining a pose estimation model with a human semantic segmentation model, sampling begins when the person is fully present and ends when they partially disappear, dynamically extracting highly relevant image frame sequences. This avoids redundant interference from traditional fixed frame rate sampling and ensures the integrity of behavioral expression.

[0031] 2. By integrating human dynamic behavior features, appearance features, and location information, a joint feature representation for cross-camera identity matching is constructed. Based on spatiotemporal constraints and a two-way matching consistency mechanism, accurate identity association of the same person under different cameras is achieved, and a continuous cross-view behavioral trajectory sequence is constructed, solving the problem of cross-view identity fragmentation in existing technologies.

[0032] 3. Personnel behavior trajectories are encoded as sequences containing location, action, and time. These sequences are then matched against preset normal templates in multiple dimensions. Confidence aggregation values ​​for abnormal behavior fragments are also incorporated to comprehensively calculate the final deviation score. This approach overcomes the limitations of traditional single-action classification, enabling deep semantic analysis of behavioral intent and trajectory deviation patterns, thereby improving the accuracy and interpretability of abnormal behavior identification. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 The flowchart shows the overall structure of a method for identifying abnormal behavior of individuals by integrating multi-source surveillance images. Detailed Implementation

[0035] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0036] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0037] In the field of smart cities and public safety, abnormal behavior recognition based on video surveillance is a core technology for preventing security risks. Existing methods generally rely on single-view video image models, which are insufficient to address the multiple challenges in complex scenarios. Firstly, to solve the problems of information limitations and video data redundancy in single-view scenarios and to achieve complete capture of multi-dimensional behavioral characteristics, a multi-camera collaborative acquisition and intelligent frame sampling scheme is adopted. Video data acquisition: Acquire raw video data from multiple surveillance cameras deployed within the target monitoring area. The video data includes image streams from different perspectives and time periods. Specifically, multi-angle cameras are deployed in the target monitoring area to simultaneously acquire raw video streams from different perspectives and time periods, covering multiple scene actions such as movement, stillness, and turning.

[0038] Then, video image frame extraction is performed: inter-frame sampling is performed on the original video stream to extract the image frame sequence, and the timestamp, camera number, and location information corresponding to each frame are retained. The inter-frame sampling operation on the original video stream first determines the start of inter-frame sampling by determining the complete appearance of the person in the video stream, and determines the end of inter-frame sampling by determining partial disappearance. The determination method for complete appearance and partial disappearance is as follows: Key point network acquisition: Using a pose estimation model, key human nodes are acquired in real time; Preliminary complete appearance judgment: When all acquired key nodes are located within the effective area of ​​the image, and the confidence of the key points is higher than the set confidence threshold, it is determined as preliminary complete appearance; Preliminary partial disappearance judgment: Define the importance weight of key nodes and calculate the weighted ratio of disappearing nodes. When the disappearance ratio exceeds the set threshold, it is determined as preliminary partial disappearance; Enhanced judgment effect: Using a human semantic segmentation model, the pixel area of ​​the human body is segmented, and the coverage rate of the human body mask and the image area is calculated. When the coverage rate is ≥90%, the preliminary complete appearance is confirmed as complete appearance; when the coverage rate is ≤40%, the preliminary partial disappearance is confirmed as partial disappearance. Specifically, the pose estimation model scans image frames in real time to locate key human body nodes and outputs the two-dimensional coordinates and confidence scores of each node. For example, when a person is standing face-up, the model can completely capture the head, torso, and limb nodes; if the person is occluded from the side, the confidence scores of some nodes may be lower than the threshold. Then, it is determined whether all key nodes are located within the effective area of ​​the image. A confidence threshold is set, and only when the confidence scores of all nodes exceed the threshold is it determined to be initially fully present. The determination of initial partial disappearance is as follows: First, importance weights are assigned to key nodes (e.g., head 0.3, torso 0.2, upper limbs 0.15, lower limbs 0.15), with core nodes (e.g., head, torso) having higher weights, reflecting their influence on behavior recognition. Then, the disappearance ratio is calculated by summing the weights of undetected nodes in the current frame and dividing by the total weight to obtain the weighted ratio of the disappeared nodes. When the ratio exceeds the set threshold, it is determined to be initially partially disappeared. Next, reinforcement judgment is performed by generating a human body mask through a human semantic segmentation model and calculating the coverage of the mask pixels with the image region. Full Presence Confirmation: If the coverage rate is ≥90%, it indicates that the human body is fully present in the frame, assisting in confirming the initial full presence as full presence and eliminating misjudgments due to occlusion in pose estimation. Partial Disappearance Confirmation: If the coverage rate is ≤40%, it indicates that the human body has mostly left the frame, assisting in confirming the initial partial disappearance as partial disappearance and avoiding misjudgments due to individual node detection errors. This step, by combining a pose estimation model with a semantic segmentation model, achieves accurate judgment of the full presence and partial disappearance states of people in video image frames, which has significant technical advantages. First, compared with traditional fixed frame rate sampling, this method starts with the person fully entering the frame and ends with partial exit, which can effectively avoid redundant interference from irrelevant frames and ensure the integrity and high relevance of the behavioral expression of the extracted frame sequence.Secondly, by judging the confidence and distribution range of key points, combined with the design of importance weights for key nodes, the determination of initial complete appearance and initial partial disappearance becomes more structural and semantic, effectively improving the ability to capture key action moments. Simultaneously, a semantic segmentation mask-assisted confirmation mechanism is introduced to correct scenarios where the pose estimation model may misjudge due to occlusion or blurring, enhancing the stability and robustness of the judgment. Furthermore, this method considers information from both static structure (key points) and image region (mask pixels) dimensions, achieving multi-feature collaborative verification, improving the accuracy of detecting the start and end of human behavior, and providing a high-quality data foundation for subsequent behavior feature extraction and trajectory construction, demonstrating good practicality and engineering adaptability.

[0039] Considering the shortcomings of existing technologies, such as insufficient accuracy in low-light environment recognition and limited behavioral features, this invention employs an image enhancement and dynamic / static feature fusion scheme for image preprocessing and feature extraction. Preprocessing is performed on the image frame sequence, and behavioral features of the human target within the sequence are extracted. Image enhancement processing is applied to the image frames, including image contrast enhancement, illumination normalization, and edge denoising, to improve the recognition accuracy of human regions in low-light environments. Static pose feature vectors are constructed using key human node information from each frame. Furthermore, dynamic behavioral features of the human body are extracted by combining the changing trends of key points between adjacent image frames to describe the human's action state in that frame. Specifically, in the image preprocessing and feature extraction steps, image enhancement processing is first applied to the extracted image frame sequence to adapt to the varying image quality in complex environments, thereby improving the stability and accuracy of subsequent recognition. Specifically, this involves three operations: First, image contrast enhancement is performed by using an adaptive histogram equalization algorithm to enhance details in dark areas of the image, making human body edges clearer. Second, illumination normalization is performed by using the Retinex algorithm to compensate for image illumination, solving the difficulty in human body recognition caused by strong backlighting or local overexposure. Third, edge-preserving noise reduction is performed by using bilateral filtering to denoise the image, smoothing image texture while preserving boundary information to facilitate subsequent target segmentation and keypoint detection. After image enhancement, the human pose feature construction stage begins. For each frame, a pre-trained human pose estimation model is invoked to identify multiple key nodes of the human body, obtain the two-dimensional coordinates and confidence values ​​of the key points, and encode them into static pose feature vectors in a preset order, representing the structural pose of the human body in the current frame. Furthermore, to introduce temporal dimension information, the position information of key points in the current frame and several adjacent frames is differentially or velocity-calculated to extract their motion trend and action continuity features, forming dynamic behavior feature vectors to characterize the action state of the human body in the current frame image. This method comprehensively considers image quality enhancement and temporal behavior modeling, and is suitable for actual monitoring environments such as low light, occlusion, and complex backgrounds. It has strong stability and practicality.

[0040] Then, to address the issue of fragmented person identification in cross-camera scenarios and achieve continuous tracking of behavior trajectories across the entire area, a multi-feature fusion cross-view matching scheme is adopted, namely, person trajectory reconstruction and identity association: The same person identified under different cameras is spatiotemporally matched to construct a cross-view person behavior trajectory sequence. Specific operations include: utilizing human dynamic behavior features and fusing appearance features and location information to form a joint feature representation for cross-camera identity matching; in multiple camera image frame sequences, the identities of people appearing under different cameras are associated based on the joint feature representation to determine whether they belong to the same target individual; based on successful identity association, a person behavior trajectory sequence is constructed according to the temporal order, spatial location, and behavioral actions of the person appearing under different cameras. Specifically, firstly, for the human target detected in each frame, its dynamic behavior features are extracted. These features consist of the displacement, angle changes, and motion trajectory of key points of human posture in adjacent image frames, used to describe the individual's short-term movement pattern. Simultaneously, its static appearance features are extracted, and the human body region image is feature-encoded using a convolutional neural network to capture information such as color, texture, and clothing, representing the distinguishable appearance features of the person. In addition, location information features, including camera ID, image frame timestamps, and the target's relative position in the image, are extracted to provide spatiotemporal context support. These three types of features are concatenated according to a unified structure to form a joint feature vector, which represents the state of a person appearing in a particular camera at each moment. Next, in multiple camera image frame sequences, all appearing person targets are matched based on the joint feature representation. By calculating cosine similarity or Euclidean distance, and combining temporal reachability constraints and spatial proximity constraints, it is determined whether they belong to the same individual. Specifically, when the joint feature similarity of the human bodies identified by two cameras is higher than a set threshold, and the time difference between the two image frames and the physical reachable path between the two cameras are reasonable, the system associates them as the same person. To prevent false matching, a bidirectional matching consistency mechanism is introduced, requiring A to match B while B also matches A for a valid match. After identity association is completed, the target personnel are sorted according to the time sequence of their appearance in each camera, and combined with their geographical location and action type information in each monitoring area, a continuous sequence of personnel behavior trajectories is constructed. This trajectory not only includes location and time, but also embeds behavioral state information, providing a semantically rich path data foundation for subsequent abnormal behavior analysis, and has good temporal continuity and spatial traceability.

[0041] Finally, to address the limitations of traditional single-action classification and achieve a deeper understanding of behavioral intent, a behavioral rule deviation analysis method is employed to calculate abnormal behavior deviation scores. The individual's behavioral trajectory is matched against a preset template to calculate the abnormal behavior deviation score. If the deviation score exceeds a set threshold, the individual is identified as exhibiting abnormal behavior. The specific steps for matching the individual's behavioral trajectory against the preset template to calculate the abnormal behavior deviation score include: encoding the individual's behavioral trajectory into a time-ordered behavioral sequence, which consists of location nodes, behavioral actions, and behavioral time; selecting the normal behavior template sequence that most closely matches the individual's behavior type and path structure; and calculating the trajectory matching score S based on time consistency, location path similarity, and the degree of action sequence matching. match The calculation method for time consistency is as follows: Where N is the trajectory length, T t T represents the timestamp at time t. t norm The timestamp represents the template time t; the method for calculating location path similarity is as follows: Where M represents the total number of frames in the normal template, L t This represents the positional characteristics at time t. The positional features at time t of the template are represented; the calculation method for the degree of action sequence matching is as follows: Where K represents the number of critical action points, and I() is an indicator function that returns 1 if the condition is met and 0 otherwise. The behavioral characteristics at time t are represented by these characteristics. The template time t represents the behavioral action features; the three are weighted and fused to obtain the trajectory matching score. Multiple potential abnormal behavior segments are identified in the behavioral trajectory, and each abnormal behavior segment is assigned an abnormal behavior confidence level C. j The aggregate confidence value S for anomalous behavior is calculated by combining the duration. abn The confidence aggregation value S abn The calculation method is as follows: Where J represents the number of potential anomalous behavior segments identified. Finally, the trajectory matching score and the anomaly aggregation value are combined to construct the final deviation score S. deviation The calculation method is as follows: S deviation =(1-S match )·S abn Finally, a judgment operation is performed. If the deviation score of the abnormal behavior exceeds the set threshold, the person is identified as having abnormal behavior.

[0042] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying individuals exhibiting abnormal behavior by integrating multi-source surveillance images, characterized in that, Includes the following steps: S1. Video Data Acquisition: Acquire raw video data from multiple surveillance cameras deployed within the target monitoring area. The video data includes image streams from different perspectives and time periods. S2. Video Image Frame Extraction: Perform inter-frame sampling on the original video stream, extract the image frame sequence, and retain the timestamp, camera number, and location information corresponding to each frame. S3. Image preprocessing and feature extraction: Perform preprocessing operations on the image frame sequence and extract the behavioral features of human targets in the image frame sequence; S4. Personnel Trajectory Reconstruction and Identity Association: Spatiotemporal identity matching will be performed on the same person identified under different cameras to construct a cross-view personnel behavior trajectory sequence; S5. Behavioral rule deviation analysis: Match the personnel's behavioral trajectory with the preset template, calculate the deviation score of abnormal behavior, and if the deviation score of abnormal behavior exceeds the set deviation score threshold, the personnel are identified as abnormal behavior personnel. The specific operations of step S5, which matches the personnel's behavioral trajectory with the preset template and calculates the deviation score of abnormal behavior, include: The behavioral trajectory of a person is encoded into a time-ordered behavioral sequence, which consists of location nodes, behavioral actions, and behavioral times. The normal behavior template sequence that is closest to the individual's behavior type and path structure is selected. Based on temporal consistency, location path similarity, and action sequence matching degree, the trajectory matching score is calculated. The calculation method for time consistency is as follows: Where N is the trajectory length, The timestamp representing time t. The timestamp represents the template time t; the method for calculating location path similarity is as follows: Where M represents the total number of frames in the normal template. This represents the positional characteristics at time t. The positional features at time t of the template are represented; the calculation method for the degree of action sequence matching is as follows: Where K represents the number of key action points, This is an indicator function that returns 1 if the condition is met, and 0 otherwise. The behavioral characteristics at time t are represented by these characteristics. The template time t represents the behavioral and action features; the trajectory matching score is obtained by weighted fusion of the three. Multiple potential anomalous behavior segments are identified in the behavioral trajectory, and each anomalous behavior segment is assigned an anomalous behavior confidence level. Calculate the aggregate confidence value of anomalous behavior by combining the duration. The confidence aggregation value The calculation method is as follows: , where J represents the number of potential anomalous behavior segments identified; Finally, the trajectory matching score and the anomaly aggregation value are combined to construct the final deviation score. The calculation method is as follows: .

2. The method for identifying abnormal behavior personnel by integrating multi-source surveillance images according to claim 1, characterized in that, In step S2, the operation of inter-frame sampling of the original video stream is to first determine the complete appearance of the person in the video stream as the start of inter-frame sampling, and determine the partial disappearance as the end of inter-frame sampling.

3. The method for identifying abnormal behavior personnel by integrating multi-source surveillance images according to claim 2, characterized in that, The method for determining whether something is completely present or partially absent is as follows: Key point network acquisition: Employs a pose estimation model to acquire key human body nodes in real time; Preliminary complete appearance judgment: When all the collected key nodes are located within the effective area of ​​the image, and the confidence of the key points is higher than the set confidence threshold, it is judged as preliminary complete appearance; Preliminary partial disappearance judgment: Define the importance weight of key nodes and calculate the weighted ratio of disappearing nodes. When the disappearance ratio exceeds the set threshold, it is judged as preliminary partial disappearance. Enhanced judgment effect: A human semantic segmentation model is used to segment the pixel region of the human body and calculate the coverage of the human body mask and the image region. When the coverage is ≥90%, the auxiliary confirmation will judge the initial complete appearance as complete appearance; when the coverage is ≤40%, the auxiliary confirmation will judge the initial partial disappearance as partial disappearance.

4. The method for identifying abnormal behavior personnel by integrating multi-source surveillance images according to claim 1, characterized in that, The image preprocessing and feature extraction operation in step S3 includes: Image enhancement processing is performed on the image frame, including image contrast enhancement, illumination normalization processing and edge denoising processing, to improve the recognition accuracy of human body regions in low-light environments. A static pose feature vector is constructed using the key human node information of each frame image. Furthermore, by combining the changing trends of key points between adjacent image frames, dynamic behavioral features of the human body are extracted to describe the human body's action state in the image frame.

5. The method for identifying abnormal behavior personnel by integrating multi-source surveillance images according to claim 1, characterized in that, The steps S4, personnel trajectory reconstruction and identity association operations, include: By utilizing human dynamic behavior characteristics and fusing appearance features with location information, a joint feature representation for cross-camera identity matching is formed. In multiple camera image frame sequences, the identities of people appearing under different cameras are associated based on joint feature representation to determine whether they belong to the same target individual; Based on successful identity verification, a sequence of personnel behavior trajectories is constructed according to the time sequence, spatial location, and actions of personnel appearing under different cameras.

Images