A cross-camera identification tracking method, device, equipment and storage medium

Abstract

The application discloses a cross-camera identification tracking method and device, equipment and a storage medium, and belongs to the technical field of computer vision and artificial intelligence. The method comprises the following steps: when it is detected that a target object meets a preset area switching condition, first feature information of the target object is extracted, and at least one target adjacent monitoring area where the target object is likely to enter and a predicted arrival time thereof are predicted; within the predicted arrival time, second feature information of a newly appeared object in the target adjacent monitoring area is acquired; similarity calculation is performed on the second feature information and the first feature information, and a comprehensive confidence is obtained by correcting a similarity calculation result in combination with space-time constraint information; and if the comprehensive confidence meets a preset condition, ID inheritance is performed. Through the application, the technical problems of high identification delay, low identification accuracy and easy loss of ID in the cross-camera tracking process in the prior art are solved.

Description

Technical Field

[0001] This application belongs to the field of computer vision and artificial intelligence technology, and specifically relates to a cross-camera recognition and tracking method, device, equipment and storage medium. Background Technology

[0002] Currently, multi-camera monitoring systems have been widely used in various scenarios such as urban fishery waters, roads, and parks. Their core requirement is to achieve continuous tracking of target objects across different camera views, ensuring the consistency of target object IDs, and providing reliable support for subsequent behavior analysis, target object tracing, and risk warning.

[0003] In existing technologies, cross-camera tracking is mainly achieved by "re-identification (ReID) technology as the core, combined with single-camera tracking algorithms". The specific technical path is as follows: First, within each single camera's field of view, target objects are selected using object detection algorithms (such as YOLO series, Faster R-CNN, etc.). Inter-frame continuous tracking of the target objects is achieved using single-camera tracking algorithms (such as ByteTrack, DeepSORT, etc.), and a temporary ID is assigned to each target object. Second, the appearance feature vectors of the target objects are extracted, and ReID technology is used to match the features of the target objects under different cameras. When a target object crosses the field of view of a camera, the ID of the target object from the previous camera is inherited to the next camera, completing cross-camera tracking.

[0004] While ReID technology has achieved cross-camera tracking of target objects to some extent, it still has many technical shortcomings in real-world complex scenarios, making it difficult to meet the requirements for high-precision, low-latency tracking. Specific drawbacks are as follows: 1. Huge computational resource consumption and poor real-time performance: Most existing ReID technologies rely on brute-force search of global features, which consumes huge computational resources. In large-scale camera networks, the recognition latency is too high, which cannot meet the real-time tracking requirements. 2. Low recognition accuracy, prone to mismatch and ID jump: When the target objects have similar features (such as people wearing uniforms, fishing boats or vehicles with similar appearances), the ReID recognition accuracy based solely on appearance features will drop significantly, and problems such as mismatch and ID jump will easily occur. 3. Poor tracking continuity and easy ID loss: Existing technologies lack an effective active relay mechanism and mostly adopt the mode of "passively waiting for the target object to appear before matching". They do not make full use of the spatial topological relationship between cameras and the temporal correlation information of the target object's movement. As a result, when the target object switches between cameras, it is easy to lose its ID due to factors such as occlusion, changes in lighting, and differences in viewing angle, resulting in poor tracking continuity. 4. Poor adaptability and difficulty in handling complex scenarios: Existing technologies have not been optimized for the specific characteristics of different application scenarios. For example, in the smart fishery administration scenario, low light conditions (including water surface reflection and wave interference) will further reduce the accuracy of feature extraction, resulting in extremely poor tracking performance of existing solutions in such scenarios, which cannot meet the actual application needs.

[0005] Therefore, there is an urgent need for a technical solution that can achieve high-precision, low-latency cross-camera tracking of target objects in complex surveillance networks. Summary of the Invention

[0006] This application provides a cross-camera recognition and tracking method, apparatus, device, and storage medium, which can solve the technical problems of high recognition latency, low recognition accuracy, easy ID loss, and poor adaptability in the cross-camera tracking process in the prior art.

[0007] To achieve the above objectives, this application provides the following technical solution: A cross-camera identification and tracking method, the method comprising: When a target object is detected to meet a preset area switching condition, the first feature information of the target object is extracted, and at least one adjacent monitoring area that the target object may enter and its estimated arrival time are predicted. Within the estimated arrival time, the second feature information of a newly appearing object in the adjacent monitoring area is obtained. The similarity between the second feature information and the first feature information is calculated, and the similarity calculation result is corrected by combining spatiotemporal constraint information to obtain a comprehensive confidence score. If the comprehensive confidence score meets a preset condition, the global identifier of the target object in the current monitoring area is assigned to the corresponding matching object in the adjacent monitoring area.

[0008] Based on the same inventive concept, embodiments of this application also provide a cross-camera identification and tracking device, the device comprising: The first information extraction module is configured to extract the first feature information of the target object when the target object is detected to meet the preset area switching conditions, and predict at least one target adjacent monitoring area that the target object may enter and its expected arrival time. The second information extraction module is configured to acquire second feature information of newly appearing objects within the adjacent monitoring area of ​​the target within the estimated arrival time. The calculation module is configured to perform similarity calculation between the second feature information and the first feature information, and to correct the similarity calculation result by combining spatiotemporal constraint information to obtain a comprehensive confidence level. The cross-camera recognition and tracking module is configured to assign the global identifier of the target object in the current monitoring area to the corresponding matching object in the adjacent monitoring area if the overall confidence level meets the preset conditions.

[0009] Based on the same inventive concept, this application also provides an electronic device, including: a memory and a processor; the processor is used to read and execute a computer program stored in the memory to implement the steps of the aforementioned cross-camera recognition and tracking method.

[0010] Based on the same inventive concept, embodiments of this application also provide a computer storage medium storing computer-executable instructions, which, when executed, implement the steps of the aforementioned cross-camera recognition and tracking method.

[0011] Compared with the prior art, this application has the following advantages: By extracting multi-dimensional feature information, the robustness of the re-identification algorithm in complex environments is enhanced, and it can effectively cope with complex environments such as low light, reflection, and occlusion, making it more adaptable. By predicting the target's adjacent cameras and their expected arrival times, there is no need to perform a brute-force search on all features. Instead, local feature matching is performed only on the predicted target's adjacent cameras, which significantly reduces the consumption of computing resources. At the same time, the target's adjacent cameras are woken up in advance to ensure timely target capture. This low-latency relay tracking improves the real-time performance of recognition, effectively enhances the system's response speed, and meets the requirements of real-time tracking. By employing an ID inheritance mechanism and an active relay strategy, seamless switching of the target object across multiple cameras is achieved. On the one hand, this avoids ID jumps when the target object crosses cameras, ensuring the continuity of the global tracking trajectory and providing reliable support for subsequent target object tracing and behavior analysis. On the other hand, combined with a spatiotemporal constraint model, it effectively filters out mismatches caused by occlusion, lighting changes, viewing angle differences, and similar target features, significantly improving tracking accuracy.

[0012] This application solves the technical problems of high recognition latency, low recognition accuracy, easy ID loss, and poor adaptability in the cross-camera tracking process of the prior art.

[0013] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the overall architecture of an embodiment of this application; Figure 2 A flowchart illustrating the method provided in the embodiments of this application; Figure 3 This is a schematic diagram of the camera topology and target object transfer prediction according to an embodiment of this application; Figure 4 This is a schematic diagram of a multi-dimensional feature vector extraction process according to an embodiment of this application; Figure 5 This is a schematic diagram of the functional modules of an embodiment of the cross-camera recognition and tracking device of this application; Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] To address the shortcomings of existing technologies, refer to Figure 1 , Figure 1 This is a schematic diagram of the overall architecture of an embodiment of this application. Figure 1 As shown, this application includes: multiple video acquisition terminals, an edge processing module, a central server, a storage module, and a visualization display terminal.

[0018] Multiple video acquisition terminals: As the front-end sensing unit of the system, they are used to acquire real-time video streams of the monitored area, transmit the video data to the edge processing module in real time, and receive high-frequency recognition, sleep and other control commands issued by the edge processing module.

[0019] Edge processing module: Deployed on each video acquisition terminal, it possesses independent real-time computing capabilities. Its core functions include single-camera target detection, multi-dimensional feature extraction, relay trigger status determination, and receiving and responding to scheduling instructions from the central server. The edge processing module has three core processing capabilities: Single-camera target detection: Real-time target (fishing boat, personnel) detection in the front-end transmitted video, locking onto and tracking objects within the frame; Feature extraction and encoding: Extracting multi-dimensional features of the target's appearance, posture, and attachments, and completing standardized encoding; Local trajectory maintenance: Recording local motion data such as the target's trajectory, displacement vector, and direction of travel within the camera's frame. After completing local preprocessing, the edge module aggregates target features, trajectory data, and relay trigger signals and uploads them to the central server, offloading a significant amount of basic detection computing power to the edge and reducing the central server's computing load.

[0020] The central server, as the core scheduling unit of the system, establishes bidirectional communication connections with each edge processing module. Its main functions include maintaining the global feature library, managing the camera topology map, executing cross-camera relay scheduling logic, and handling global tasks such as target loss warnings and global ID allocation. The central server adopts a distributed architecture, supporting massive camera access and high-concurrency feature comparison requests. As the scheduling hub of the entire system, it integrates four core software functions: a cross-camera spatiotemporal correlation engine (combining target motion parameters and spatial location to calculate target cross-camera travel time and spatial constraints); a global feature library (summarizing target feature codes uploaded from all edges and uniformly storing target feature files for the entire scene); topology management (maintaining the topology map of all camera points, calculating the target transfer probability between adjacent cameras, and constructing a weighted directed topology network); and relay scheduling logic (receiving an edge relay trigger command, predicting downstream candidate cameras based on topology and spatiotemporal data, issuing warnings and feature scheduling commands to achieve active cross-camera relay tracking). The central server connects downwards to the storage module and visualization terminal, and upwards to coordinate all edge nodes.

[0021] Storage module: Communicates with the central server to persistently store key data such as target trajectory data, multi-dimensional feature information, and relay matching records, providing data support for subsequent target tracing and system optimization; Visualization terminal: It communicates with the central server and is used to display information such as the target's cross-camera relay tracking trajectory, target ID, and matching confidence level in real time, supporting real-time monitoring and operation by management personnel.

[0022] By leveraging edge computing to process high-frequency image tasks, the central cloud handles global scheduling, balancing response speed and global consistency. The edge devices process high-frequency tasks such as target detection and feature extraction locally, reducing data transmission latency and improving real-time response capabilities. The central cloud manages camera topology, global ID allocation, and trajectory stitching, ensuring global consistency in target tracking across multi-camera networks and adapting to the application needs of large-scale monitoring scenarios.

[0023] Reference Figure 2 This application discloses a cross-camera recognition and tracking method, the method comprising: Step S10: When the target object is detected to meet the preset area switching conditions, the first feature information of the target object is extracted, and the target object may enter at least one adjacent monitoring area and its expected arrival time is predicted. In this embodiment, within the monitoring area of ​​the current camera, a deep learning algorithm is used to perform real-time detection and single-camera tracking of the target object, and to extract multi-dimensional feature vectors of the target object, including visual features, posture features and features of the objects it carries, in real time. The preset area switching condition is that the target object enters the preset edge trigger area within the current camera's monitoring area, and the direction of its motion vector points towards the blind spot. The system monitors the target object's position in the current camera's view in real time. When the target object enters the preset edge trigger area within the current camera's monitoring area, and the direction of its motion vector points towards the blind spot, it is determined to be in a relay trigger state. Based on the target object's motion vector, the current camera's geographical location information, and a pre-constructed camera topology map, the system calculates at least one adjacent camera that the target object may enter and its estimated arrival time.

[0024] Further, in one embodiment, the first feature information includes: global visual features, local pose features, carried object features, and gait spatiotemporal features; the extraction of the first feature information of the target object includes: Using convolutional neural networks or visual models, feature encoding is performed on the images of target objects within the monitoring area of ​​the current camera to extract global visual features that can characterize the overall appearance of the target objects; The algorithm detects the key skeletal node coordinates of the target object within the monitoring area of ​​the current camera by using a human key point detection algorithm, and extracts the local pose features of the target object based on the node coordinate relationship. By using an attribute recognition network, the characteristics of the objects carried by the target objects within the current camera's monitoring area are extracted; Gait spatiotemporal features are extracted from consecutive multi-frame image sequences.

[0025] In this embodiment, the first feature information includes global visual features, local pose features, carried object features, and gait spatiotemporal features. Specifically, a Convolutional Neural Network (CNN) or Visual Transformer (ViT) model is used to encode features of the target object image within the current camera's monitoring area, extracting global visual features that characterize the overall appearance of the target, covering core information such as the target's contour, texture, and overall color distribution.

[0026] By using human keypoint detection algorithms (such as HRNet and YOLO-Pose), the key skeletal nodes (such as head, torso, and limb joints) of the target object within the monitoring area of ​​the current camera are accurately detected. Based on the node coordinate relationship, the local pose features of the target object are extracted to distinguish similar appearance targets with obvious pose differences.

[0027] By utilizing attribute recognition networks, the characteristics of the objects carried by target objects (such as backpacks, hand-held items, fishing gear, etc.) within the current camera's monitoring area can be accurately identified and extracted, further enriching the dimensions of target features and improving feature uniqueness.

[0028] Reference Figure 4 , Figure 4 This is a schematic diagram illustrating the multi-dimensional feature vector extraction process according to an embodiment of this application. Figure 4 As shown, 1. First input: raw image frame. A single frame of surveillance footage captured by the camera, containing all pixel information such as background, pedestrians, and debris, is the raw input of the entire network.

[0029] 2. Locate and crop the target human body region from complex images using a human detection network. Filter out irrelevant backgrounds such as walls, floors, and shelves; output only local image bounding boxes containing pedestrians for the multi-branch feature extraction module to perform high-precision feature extraction.

[0030] 3. Multi-branch feature extraction: After detecting a human body, four independent branches are simultaneously activated to extract human body information from completely different dimensions. The four branches operate in parallel without interference. Global Visual Features (CNN / VIT) extracts the overall appearance features of the entire human body, such as clothing color, overall body shape, coat style, skin color, and other macroscopic visual information of the entire human body.

[0031] Local pose features (keypoints), namely key points of the human body (head, shoulders, elbows, hips, knees, ankles, etc.), extract the posture of human skeletal joints: walking posture, raising hands, bending over, turning sideways and other action postures, which are not slightly affected by clothing.

[0032] Bag / Phone features are specifically designed to detect and extract features of pedestrians' personal belongings, such as handbags, backpacks, mobile phones, suitcases, hats, etc., to distinguish similar pedestrians who cannot be identified by their physical appearance alone.

[0033] Gait spatiotemporal features (Sequence) are gait features extracted from a sequence of multiple consecutive frames of images (the rhythm, stride, and temporal pattern of limb swing). As a temporal feature, it captures the unique dynamic patterns of human walking and has stronger stability in cross-view recognition.

[0034] 4. Feature Fusion Layer: This layer fuses the four independent features output from the four branches above—global appearance, skeletal pose, personal items, and temporal gait—through attention. This addresses the limitations of single-feature analysis, such as the tendency to see identical clothing or insufficient differentiation based solely on pose. By integrating these four complementary features, it eliminates the shortcomings of single-branch recognition.

[0035] 5. Output: Multidimensional feature vector: After fusion, a fixed-length one-dimensional digital vector is obtained, which is the pedestrian's unique digital "identity code" and can be used for cross-camera pedestrian retrieval, trajectory matching, personnel tracking, and security missing person search.

[0036] By combining detailed features such as pose and carried objects, the robustness of the re-identification algorithm in complex environments is enhanced. Breaking away from the limitations of traditional methods that rely solely on appearance features, this algorithm integrates detailed features such as target pose and carried objects, combining them with appearance features to form a comprehensive feature representation. This effectively addresses complex environments such as low light, reflection, and occlusion, improving the accuracy of feature extraction and the reliability of matching.

[0037] Further, in one embodiment, predicting at least one adjacent monitoring area that the target object may enter and its estimated arrival time includes: Calculate the extended intersection of the motion vector of the target object in the constructed camera topology map to determine the target adjacent monitoring area of ​​at least one target adjacent camera; Based on the current geographic location information of the camera, the physical distance between the target and the adjacent cameras, and the historical average transfer time, combined with the target object's current speed, the estimated time for the target object to reach the adjacent camera is calculated.

[0038] In this embodiment, taking camera A tracking a fishing boat as an example, the edge processing module solves the navigation motion vector based on the continuous frame coordinates of the fishing boat, extends the vector along the direction of travel, and calculates the spatial intersection with the preset camera topology geographical path. Then, it filters out the adjacent camera B and determines that camera B is the target adjacent camera, and the monitoring area of ​​camera B is the target adjacent monitoring area.

[0039] The physical distance and historical travel time from camera A to camera B are obtained. The theoretical travel time is calculated by combining the real-time sailing speed of the fishing boat. The time for the target object (e.g., fishing boat) to arrive at camera B is 10-15 seconds by integrating the historical statistical intervals.

[0040] In another embodiment, taking camera A tracking a fishing boat as an example, the edge processing module solves the navigation motion vector based on the continuous frame coordinates of the fishing boat, extends the vector along the direction of travel, and calculates the spatial intersection with the preset camera topological geographical path. The intersecting adjacent cameras B and C are then selected, and cameras B and C are determined as target adjacent cameras. The monitoring areas of cameras B and C are the target adjacent monitoring areas.

[0041] The physical distance and historical travel time from camera A to camera B and from camera A to camera C are obtained. The theoretical travel time is calculated by combining the real-time sailing speed of the fishing boat. By integrating the historical statistical intervals, it is found that the time for the target object (e.g., fishing boat) to arrive at camera B is 10-15s and the time to arrive at camera C is 20-30s.

[0042] The camera topology diagram is constructed as follows: Acquire historical transfer data of the target object between different cameras; Based on the historical transfer data, the probability of target object transfer between any two cameras is obtained; A weighted directed graph is constructed with each camera as a node in the graph and the probability as the edge weight. This weighted directed graph serves as the camera topology graph.

[0043] In this embodiment, multiple video capture terminals with AI computing power are first deployed within the target area. Each terminal has a built-in edge processing module and a pre-installed deep learning inference engine. The central server connects to all terminals via a network and imports a map of the camera distribution in the area. By analyzing historical video data, it automatically learns and constructs a camera topology map, recording the average travel time and transition probability between different cameras.

[0044] Specifically, the system collects and records historical data on the transfer of a target object between different cameras over a long period of time, including key information such as the transfer path (i.e., the specific link by which the target object is transferred from the current camera to the next camera), the physical distance between the two cameras, and the actual time consumed in the transfer of the target object. Statistical methods (such as maximum likelihood estimation) are used to analyze the collected historical transfer data, calculate the probability of target object transfer between any two camera pairs, and determine the likelihood of target object transferring from the current camera to each candidate adjacent camera; Using each camera as a node in the graph and the transition probability between cameras as the edge weight, a weighted directed graph is constructed. This weighted directed graph is the camera topology graph. The larger the edge weight, the higher the probability that the target will transfer from the corresponding starting camera to the ending camera, providing accurate probability support for the prediction of subsequent adjacent cameras.

[0045] By introducing camera topology and temporal constraints, interference targets with similar visual features but inconsistent spatiotemporal logic are effectively filtered out. By verifying whether the target appears within the expected time and whether its location conforms to the camera topology reachability, false matching caused by similar target appearances is avoided, significantly improving the accuracy of cross-camera matching and reducing ID jump phenomena. Step S20: Within the estimated arrival time, acquire the second characteristic information of newly appearing objects within the adjacent monitoring area of ​​the target; In this embodiment, a relay warning instruction is sent to each adjacent camera of the target. The relay warning instruction includes the first feature information of the target and the estimated arrival time. Each adjacent camera to the target activates a high-frequency recognition mode within the estimated arrival time to extract multi-dimensional features from the newly appearing object, obtaining second feature information of the newly appearing object within the target monitoring area of ​​the adjacent camera. This second feature information includes global visual features, local pose features, carried object features, and gait spatiotemporal features.

[0046] By employing an active relay mechanism, adjacent cameras are activated in advance to avoid missing capture of targets when they cross camera boundaries. Combined with a spatiotemporal constraint model, it effectively filters out false matches caused by occlusion, changes in lighting, differences in viewing angles, and similarity in target features, thus significantly improving tracking accuracy.

[0047] Step S30: Calculate the similarity between the second feature information and the first feature information, and correct the similarity calculation result by combining the spatiotemporal constraint information to obtain the comprehensive confidence level; In this embodiment, the first and second feature information are first split into four sub-feature vectors: global vision, local pose, carried object, and gait spatiotemporal. The cosine similarity is used to calculate the sub-item similarity of each sub-feature. The weighting coefficients of the four types of features are adaptively configured and normalized according to the current acquisition environment. The similarity calculation result is obtained by weighted summation.

[0048] Furthermore, in one embodiment, the step of correcting the similarity calculation result by incorporating spatiotemporal constraint information to obtain a comprehensive confidence score includes: Calculate the time difference between the first target time point and the second target time point, and use the time difference as the actual duration of the target transfer; wherein, the first target time point is the time point when the target object disappears from the current camera monitoring area, and the second target time point is the time point when the target object appears in the monitoring area of ​​the adjacent camera. The detection result is obtained by checking whether the actual duration is within the expected arrival time range; Based on the detection results, the time offset is determined; Based on the aforementioned time offset, the similarity calculation results are corrected to obtain the overall confidence level.

[0049] In this embodiment, the first target time point when the target object disappears from the current camera monitoring area and the second target time point when the target object appears in the monitoring area of ​​the adjacent camera are accurately recorded. The time difference between the two time points is calculated, which is the actual time of the target transfer. The calculated actual duration is compared with the expected arrival time range [T_min, T_max] to check if the actual duration is within the expected arrival time range. If the actual duration is within the expected arrival time range, the deviation value (i.e., time offset) between the actual duration and the time is positive. If the actual duration exceeds the time range, the deviation value (i.e., time offset) between the actual duration and the time is negative, and the larger the absolute value of the time offset, the larger the time deviation.

[0050] Penalize or reward similarity calculation results based on time offset: When the time offset is positive and within the preset range, a positive reward is given to the similarity calculation result to improve the similarity score, resulting in a corrected similarity score, i.e., the overall confidence level. When the time offset is negative or exceeds the preset range, the similarity calculation result is penalized in reverse to reduce the similarity score, resulting in a corrected similarity score, i.e., the overall confidence level.

[0051] Further, in one embodiment, after step S30, the following is included: If the overall confidence level does not meet the preset conditions, an alarm signal indicating that the target object is lost is sent to the central server. The alarm signal indicating the loss of the target object includes the first characteristic information of the target object, its last location coordinates, and the expected relocation range.

[0052] In this embodiment, if there is no overall confidence level greater than or equal to a preset threshold, an alarm signal indicating that the target object is lost is sent to the central server.

[0053] The alarm signal includes the target object's first characteristic information, its last location coordinates, and its expected transfer range, which facilitates overall scheduling by the central server and improves the probability of recapture of the target object.

[0054] Step S50: If the overall confidence level meets the preset conditions, then the global identifier of the target object in the current monitoring area is assigned to the corresponding matching object in the adjacent monitoring area of ​​the target.

[0055] In this embodiment, the preset condition is that the overall confidence score is greater than or equal to a preset threshold. If there is a corrected similarity score (overall confidence score) that is greater than or equal to the preset threshold, the global ID of the target object in the current camera's monitoring area is assigned to the newly entered target object in the target monitoring area of ​​the target adjacent camera corresponding to the highest overall confidence score, thus completing cross-camera tracking.

[0056] Furthermore, referring to Figure 3 , Figure 3 This is a schematic diagram illustrating the camera topology and target object transfer prediction according to an embodiment of this application. Figure 3 As shown, the circles represent the physically deployed camera locations: cameras A, B, C, and D form a chain topology. Camera A is the starting point for capturing the target. After the target leaves A, there are two paths: towards camera B or towards camera C. Cameras B and C are intermediate nodes, and both can eventually connect to the end camera D, forming two complete cross-camera links: Link ①: A→B→D; Link ②: A→C→D.

[0057] Each link is labeled with the estimated arrival time of the target object from the starting point to the next station, used for spatiotemporal constraint verification and prediction of the target object's arrival time period: Path 1 (A→B): Target object travel time 10~15s; Path 3 (B→D): Target object travel time 5~8s; Path 2 (A→C): Target object travel time 20~30s; Path 4 (C→D): Target object travel time 12~18s; During central dispatch, based on the target object's appearance time and path duration at camera A, the estimated arrival time of the target object at cameras B / C / D is calculated in advance, and a time-limited high-frequency capture warning is issued to downstream cameras.

[0058] Based on historical big data statistics, the probability of traffic diversion for the target object corresponding to camera A is obtained: P_AB=0.6, meaning that after the target object leaves the camera A, there is a 60% probability that it will move towards the camera B; P_AC=0.4, meaning that after the target object leaves the camera A, there is a 40% probability that it will move towards the camera C.

[0059] The specific details are as follows, in conjunction with this application: When a target enters the field of view of camera A, the edge processing module immediately locks onto the target and assigns a temporary ID. The system extracts the target's visual features (such as clothing color and body shape) and non-visual features (such as gait and backpack) in real time. When the target moves towards the right edge of camera A and its velocity vector points towards camera B, the system determines that a relay is triggered.

[0060] Camera A sends an alert to the central server, which then notifies Camera B based on the topology map: "Target ID-001 is expected to enter your field of view in 15-20 seconds."

[0061] After receiving the instruction, camera B focuses its scan on the left entrance area. When the target appears, camera B first compares its features with ID-001. Considering the time difference (17 seconds, as expected), the system confirms with 98% confidence that the target is ID-001, completing the relay.

[0062] This application is applicable to a variety of monitoring scenarios (such as smart fishery administration, park security, traffic monitoring, etc.), and is especially suitable for scenarios with similar target characteristics and complex environments (such as low light, reflection, and obstruction). It has strong practicality and promotional value.

[0063] In this embodiment, when a target object is detected to meet a preset area switching condition, the first feature information of the target object is extracted, and at least one adjacent monitoring area that the target object may enter and its estimated arrival time are predicted; within the estimated arrival time, the second feature information of a newly appearing object in the adjacent monitoring area is obtained; the second feature information and the first feature information are used to calculate the similarity, and the similarity calculation result is corrected by combining spatiotemporal constraint information to obtain a comprehensive confidence level; if the comprehensive confidence level meets the preset condition, the global identifier of the target object in the current monitoring area is assigned to the corresponding matching object in the adjacent monitoring area.

[0064] By extracting multi-dimensional feature information, the robustness of the re-identification algorithm in complex environments is enhanced, and it can effectively cope with complex environments such as low light, reflection, and occlusion, making it more adaptable. By predicting the target's adjacent cameras and their expected arrival times, there is no need to perform a brute-force search on all features. Instead, local feature matching is performed only on the predicted target's adjacent cameras, which significantly reduces the consumption of computing resources. At the same time, the target's adjacent cameras are woken up in advance to ensure timely target capture. This low-latency relay tracking improves the real-time performance of recognition, effectively enhances the system's response speed, and meets the requirements of real-time tracking. By employing an ID inheritance mechanism and an active relay strategy, seamless switching of the target object across multiple cameras is achieved. On the one hand, this avoids ID jumps when the target object crosses cameras, ensuring the continuity of the global tracking trajectory and providing reliable support for subsequent target object tracing and behavior analysis. On the other hand, combined with a spatiotemporal constraint model, it effectively filters out mismatches caused by occlusion, lighting changes, viewing angle differences, and similar target features, significantly improving tracking accuracy.

[0065] This embodiment solves the technical problems of high recognition latency, low recognition accuracy, easy ID loss, and poor adaptability in the cross-camera tracking process of the prior art.

[0066] Based on the same inventive concept, embodiments of this application also provide a cross-camera identification and tracking device.

[0067] In one embodiment, reference is made to Figure 5 , Figure 5 This is a functional module diagram of an embodiment of the cross-camera recognition and tracking device of this application. Figure 5 As shown, the cross-camera recognition and tracking device includes: The first information extraction module 10 is configured to extract the first feature information of the target object when the target object is detected to meet the preset area switching conditions, and predict at least one target adjacent monitoring area that the target object may enter and its expected arrival time. The second information extraction module 20 is configured to acquire second feature information of newly appearing objects within the adjacent monitoring area of ​​the target within the estimated arrival time. The calculation module 30 is configured to perform similarity calculation between the second feature information and the first feature information, and to correct the similarity calculation result by combining spatiotemporal constraint information to obtain a comprehensive confidence level. The cross-camera recognition and tracking module 40 is configured to assign the global identifier of the target object in the current monitoring area to the corresponding matching object in the adjacent monitoring area of ​​the target if the comprehensive confidence level meets the preset conditions.

[0068] Optionally, in one embodiment, the first feature information includes: global visual features, local pose features, carried object features, and gait spatiotemporal features; the first information extraction module 10 is configured to: Using convolutional neural networks or visual models, feature encoding is performed on the images of target objects within the monitoring area of ​​the current camera to extract global visual features that can characterize the overall appearance of the target objects; The algorithm detects the key skeletal node coordinates of the target object within the monitoring area of ​​the current camera by using a human key point detection algorithm, and extracts the local pose features of the target object based on the node coordinate relationship. By using an attribute recognition network, the characteristics of the objects carried by the target objects within the current camera's monitoring area are extracted; Gait spatiotemporal features are extracted from consecutive multi-frame image sequences.

[0069] Optionally, in one embodiment, the first information extraction module 10 is configured to: Calculate the extended intersection of the motion vector of the target object in the constructed camera topology map to determine the target adjacent monitoring area of ​​at least one target adjacent camera; Based on the current geographic location information of the camera and the physical distance or historical average transfer time between the target and adjacent cameras, combined with the current speed of the target object, the estimated time for the target object to reach the target and adjacent cameras is calculated.

[0070] Optionally, in one embodiment, the camera topology diagram is constructed as follows: Acquire historical transfer data of the target object between different cameras; Based on the historical transfer data, obtain the probability distribution of target object transfer between any two cameras; A weighted directed graph is constructed with each camera as a node in the graph and the probability distribution as the edge weights. This weighted directed graph serves as the camera topology graph.

[0071] Optionally, in one embodiment, the computing module 30 is configured to: Calculate the time difference between the first target time point and the second target time point, and use the time difference as the actual duration of the target transfer; wherein, the first target time point is the time point when the target object disappears from the current camera monitoring area, and the second target time point is the time point when the target object appears in the monitoring area of ​​the adjacent camera. The detection result is obtained by checking whether the actual duration is within the expected arrival time range; Based on the detection results, the time offset is determined; Based on the aforementioned time offset, the similarity calculation results are corrected to obtain the overall confidence level.

[0072] Optionally, in one embodiment, the computing module 30 is configured to: When the time offset is positive and within a preset range, a positive reward is given to the similarity calculation result to obtain the comprehensive confidence level; When the time offset is negative or exceeds a preset range, the similarity calculation result is penalized in reverse to obtain the comprehensive confidence level.

[0073] Optionally, in one embodiment, the cross-camera identification and tracking device further includes an alarm module configured to: If the overall confidence level does not meet the preset conditions, an alarm signal indicating that the target object is lost is sent to the central server. The alarm signal indicating the loss of the target object includes the first characteristic information of the target object, its last location coordinates, and the expected relocation range.

[0074] The functions of each module in the aforementioned cross-camera recognition and tracking device correspond to the steps in the aforementioned cross-camera recognition and tracking method embodiment, and their functions and implementation processes will not be described in detail here.

[0075] Based on the same inventive concept, embodiments of this application also provide an electronic device, the structure of which is as follows: Figure 6 As shown, it includes: a memory and a processor, wherein the processor is used to read and execute the computer program stored in the memory to implement the aforementioned cross-camera recognition and tracking method.

[0076] Based on the same inventive concept, this application also provides a computer storage medium storing computer-executable instructions, which, when executed, implement the aforementioned cross-camera recognition and tracking method.

[0077] Finally, it should be noted that while some processes described in the embodiments of this application include multiple operations or steps that appear in a specific order, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0078] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A cross-camera recognition and tracking method, characterized in that, The method includes: When a target object is detected to meet the preset area switching conditions, the first feature information of the target object is extracted, and at least one target adjacent monitoring area that the target object may enter and its estimated arrival time are predicted. Within the estimated arrival time, acquire the second characteristic information of newly appearing objects within the adjacent monitoring area of ​​the target; The similarity between the second feature information and the first feature information is calculated, and the similarity calculation result is corrected by combining spatiotemporal constraint information to obtain a comprehensive confidence level. If the overall confidence level meets the preset conditions, then the global identifier of the target object in the current monitoring area is assigned to the corresponding matching object in the adjacent monitoring area of ​​the target.

2. The method according to claim 1, characterized in that, The first feature information includes: global visual features, local pose features, carried object features, and gait spatiotemporal features. Extracting the first feature information of the target object includes: Using convolutional neural networks or visual models, feature encoding is performed on the images of target objects within the monitoring area of ​​the current camera to extract global visual features that can characterize the overall appearance of the target objects; The algorithm detects the key skeletal node coordinates of the target object within the monitoring area of ​​the current camera by using a human key point detection algorithm, and extracts the local pose features of the target object based on the node coordinate relationship. By using an attribute recognition network, the characteristics of the objects carried by the target objects within the current camera's monitoring area are extracted; Gait spatiotemporal features are extracted from consecutive multi-frame image sequences.

3. The method according to claim 1, characterized in that, The prediction of at least one adjacent monitoring area that the target object may enter and its estimated arrival time includes: Calculate the extended intersection of the motion vector of the target object in the constructed camera topology map to determine the target adjacent monitoring area of ​​at least one target adjacent camera; Based on the current geographic location information of the camera and the physical distance or historical average transfer time between the target and adjacent cameras, combined with the current speed of the target object, the estimated time for the target object to reach the target and adjacent cameras is calculated.

4. The method according to claim 3, characterized in that, The camera topology diagram is constructed as follows: Acquire historical transfer data of the target object between different cameras; Based on the historical transfer data, obtain the probability distribution of target object transfer between any two cameras; A weighted directed graph is constructed with each camera as a node in the graph and the probability distribution as the edge weights. This weighted directed graph serves as the camera topology graph.

5. The method according to claim 1, characterized in that, The step of correcting the similarity calculation results by incorporating spatiotemporal constraint information to obtain a comprehensive confidence score includes: Calculate the time difference between the first target time point and the second target time point, and use the time difference as the actual duration of the target transfer; wherein, the first target time point is the time point when the target object disappears from the current camera monitoring area, and the second target time point is the time point when the target object appears in the monitoring area of ​​the adjacent camera. The detection result is obtained by checking whether the actual duration is within the expected arrival time range; Based on the detection results, the time offset is determined; Based on the aforementioned time offset, the similarity calculation results are corrected to obtain the overall confidence level.

6. The method according to claim 5, characterized in that, The process of correcting the similarity calculation results based on the time offset to obtain a comprehensive confidence score includes: When the time offset is positive and within a preset range, a positive reward is given to the similarity calculation result to obtain the comprehensive confidence level; When the time offset is negative or exceeds a preset range, the similarity calculation result is penalized in reverse to obtain the comprehensive confidence level.

7. The method according to claim 1, characterized in that, After correcting the similarity calculation results by incorporating spatiotemporal constraint information to obtain a comprehensive confidence score, the process includes: If the overall confidence level does not meet the preset conditions, an alarm signal indicating that the target object is lost is sent to the central server. The alarm signal indicating the loss of the target object includes the first characteristic information of the target object, its last location coordinates, and the expected relocation range.

8. A cross-camera recognition and tracking device, characterized in that, The device includes: The first information extraction module is configured to extract the first feature information of the target object when the target object is detected to meet the preset area switching conditions, and predict at least one target adjacent monitoring area that the target object may enter and its expected arrival time. The second information extraction module is configured to acquire second feature information of newly appearing objects within the adjacent monitoring area of ​​the target within the estimated arrival time. The calculation module is configured to perform similarity calculation between the second feature information and the first feature information, and to correct the similarity calculation result by combining spatiotemporal constraint information to obtain a comprehensive confidence level. The cross-camera recognition and tracking module is configured to assign the global identifier of the target object in the current monitoring area to the corresponding matching object in the adjacent monitoring area if the overall confidence level meets the preset conditions.

9. An electronic device, characterized in that, include: Memory, processor; The processor is configured to read and execute the computer program stored in the memory to implement the steps of the cross-camera identification and tracking method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, implement the steps of the cross-camera identification and tracking method according to any one of claims 1-7.

Images