Method and system for dense scene person continuous tracking

Abstract

The present application relates to the technical field of video monitoring target tracking, and discloses a dense scene personnel continuous tracking method and system, the method collects multi-camera multi-view image frames to perform coordinate mapping and global splicing, extracts personnel multi-granularity fusion appearance features to construct personnel feature archives, constructs a hierarchical persistent feature storage library to store off-site personnel archives, and when personnel return, hierarchical cascaded re-identification retrieval is performed according to storage priority to complete track splicing to obtain complete continuous behavior tracks, solves the problem of identity mismatch track splitting after personnel off-site return, and improves the accuracy and storage efficiency of dense scene personnel continuous tracking.

Description

Technical Field

[0001] This invention relates to the field of video surveillance target tracking technology, and in particular to a method and system for continuous tracking of people in dense scenes. Background Technology

[0002] In factory park personnel management scenarios that include canteens and operational storage areas, continuous personnel tracking technology based on computer vision is commonly used. This technology relies on cameras deployed throughout the area to collect image data, enabling real-time identity binding and trajectory recording of moving personnel. This supports personnel movement management, operational compliance checks, and access control. Existing personnel tracking algorithms optimized for dual-scenario adaptation control memory usage by dynamically maintaining an identity feature database, ensuring real-time tracking speed within a single scenario while adapting to the tracking accuracy requirements of different scenarios.

[0003] When implementing cross-regional personnel tracking management in the industrial park, if a worker temporarily leaves the camera's full coverage area and re-enters the coverage area after a period of time, the tracking system cannot maintain its original identity binding. The root cause of this problem is that the existing dynamic management mechanism of the identity feature database in continuous personnel tracking algorithms only retains the mapping relationship between the features and identity IDs of personnel currently within the coverage area. To control memory usage, the algorithm deletes mapping entries for personnel who have been away from the coverage area for a longer period. The existing dual-scene adaptation only adjusts tracking parameters under different scenarios and does not establish an identity retention mechanism across gaps in coverage areas. This results in the same person being identified as a new entrant and assigned a new identity ID upon re-entry. This issue causes the complete behavioral trajectory of the same person to be split into multiple independent trajectories. The behavioral data of personnel across scenarios cannot be integrated, making it impossible to support cross-regional identity association and behavioral analysis. It cannot accurately calculate the actual working time and compliance of worker movement, nor can it complete the complete personnel trajectory tracing in the event of material abnormalities or safety incidents, affecting the accuracy of personnel management and anomaly investigation in the factory park. Summary of the Invention

[0004] Methods for continuous tracking of people in densely populated areas:

[0005] S1: Collect multi-view image frame sequences within the coverage area of ​​multiple cameras, perform view coordinate mapping and global stitching on the multi-view image frame sequences to obtain a global stitched image frame sequence, and perform target detection and appearance feature extraction on all personnel in the current scene based on the global stitched image frame sequence to obtain a personnel feature profile set;

[0006] Further, step S1 includes:

[0007] S11: Obtain the sequence of multi-view image frames output by the matching multi-camera acquisition unit at the same time. Each image frame carries the intrinsic parameter matrix and extrinsic parameter matrix of the corresponding camera.

[0008] S12: Based on the intrinsic and extrinsic parameter matrices of each camera, calculate the overlapping area of ​​the viewpoints and the coordinate transformation relationship between adjacent cameras, and establish a unified global coordinate system for the scene;

[0009] S13: In the scene global coordinate system, perform pixel-level registration and weighted fusion stitching on adjacent image frames with overlapping view areas in the multi-view image frame sequence to eliminate view blind spots and generate a global stitched image frame sequence covering the entire scene.

[0010] S14: Perform person target detection on each frame in the global stitched image frame sequence, and obtain the bounding box coordinates of each detected person and its position coordinates in the scene global coordinate system;

[0011] S15: Perform multi-granularity appearance feature extraction on the bounding box region of each inspected person. The multi-granularity appearance feature extraction includes the following steps:

[0012] S151: Perform face feature encoding on the face region within the bounding box region to obtain the face feature vector;

[0013] S152: Perform gait cycle analysis and gait feature encoding on the whole body region within the bounding box region to obtain the gait feature vector;

[0014] S153: Perform clothing texture and body shape contour encoding on the torso region in the bounding box region to obtain the body feature vector;

[0015] S154: Concatenate the facial feature vector, gait feature vector, and body shape feature vector according to a preset dimensional concatenation order to generate the fused appearance feature vector of the person.

[0016] S16: Assign a unique identity tracking identifier to each person being detected, and bind the identity tracking identifier with the corresponding fused appearance feature vector, the position coordinates of the current frame and the timestamp to form a person feature profile. All person feature profiles constitute a set of person feature profiles.

[0017] S2: Based on the personnel feature profile set, establish an departure attenuation weight for each personnel feature profile, and construct a hierarchical persistent feature repository. When the tracked personnel leave the full coverage area, write the corresponding personnel feature profile into the hierarchical persistent feature repository according to the departure attenuation weight to obtain the hierarchical persistent feature repository.

[0018] Further, step S2 includes:

[0019] S21: Perform inter-frame position matching for each personnel feature file in the personnel feature file set between consecutive frames, and determine whether the personnel corresponding to each personnel feature file still appears in the latest frame of the current global stitched image frame sequence;

[0020] Specifically, the inter-frame position matching method is as follows: calculate the Euclidean distance between the position coordinates of each detected person in the current frame and the position coordinates recorded in the feature files of each person in the previous frame, and at the same time calculate the cosine similarity between the fused appearance feature vector of the detected person in the current frame and the fused appearance feature vector in the feature files of each person. Pairs with Euclidean distance lower than the position matching threshold and cosine similarity higher than the appearance matching threshold are determined to be the same person.

[0021] S22: If a personnel feature profile is not matched with any detected personnel in the current frame within a consecutive preset number of frames, then the personnel feature profile is marked as leaving the field and the departure start timestamp is recorded.

[0022] S23: Calculate the departure attenuation weight for the personnel feature files marked as departing;

[0023] Specifically, the departure attenuation weight is calculated as follows: starting from the departure start time stamp, the departure attenuation weight value is gradually reduced as the departure duration increases according to the preset time attenuation function. The time attenuation function is an exponential attenuation function, and the attenuation rate is determined by the scene type parameter. The attenuation rate of the canteen scene is slower than that of the warehouse scene.

[0024] S24: Construct a hierarchical persistent feature repository, which includes three levels: hot storage area, warm storage area and cold storage area;

[0025] S241: When a personnel feature file is marked as being out of the field and the duration of the absence does not exceed the first duration threshold, the personnel feature file is written into the hot storage area. The personnel feature file in the hot storage area maintains the complete fused appearance feature vector and all historical location coordinate sequences.

[0026] S242: When the duration of absence exceeds the first duration threshold but does not exceed the second duration threshold, the personnel feature file is transferred from the hot storage area to the warm storage area. The personnel feature file in the warm storage area retains the fused appearance feature vector and the most recent preset number of historical location coordinates.

[0027] S243: When the duration of absence exceeds the second duration threshold but does not exceed the third duration threshold, the personnel feature file is transferred from the warm storage area to the cold storage area. The personnel feature file in the cold storage area only retains the fused appearance feature vector and identity tracking identifier.

[0028] S244: When the duration of absence exceeds the third duration threshold, delete the personnel's profile from the cold storage area to free up storage space;

[0029] Specifically, the first duration threshold, the second duration threshold, and the third duration threshold are configured differently according to the scenario type parameter. In the canteen scenario, all three duration thresholds are greater than the corresponding duration thresholds in the warehouse scenario.

[0030] S3: When a person enters the full coverage area, the fused appearance feature vector of the person is extracted as the feature vector to be matched. In the hierarchical persistent feature repository, the feature vector to be matched is subjected to hierarchical cascaded re-identification retrieval according to the priority order of hot storage area, warm storage area and cold storage area to obtain the identity backtracking matching result.

[0031] Further, step S3 includes:

[0032] S31: Perform the same multi-granularity appearance feature extraction as in step S15 on newly entered personnel in the full coverage area to generate the personnel's feature vector to be matched;

[0033] S32: In the hot storage area of ​​the hierarchical persistent feature repository, calculate the weighted similarity between the feature vector to be matched and the fused appearance feature vector of each personnel feature profile in the hot storage area;

[0034] Specifically, the weighted similarity is calculated as follows: the cosine similarity between the face feature vector, gait feature vector, and body shape feature vector in the feature vector to be matched and the corresponding parts in the personnel feature file is calculated separately. Then, the similarity of the face feature vector is multiplied by the face weight coefficient, the similarity of the gait feature vector is multiplied by the gait weight coefficient, and the similarity of the body shape feature vector is multiplied by the body shape weight coefficient, and the results are summed to obtain the weighted similarity. The face weight coefficient is greater than the gait weight coefficient, and the gait weight coefficient is greater than the body shape weight coefficient.

[0035] S33: Multiply the weighted similarity by the departure attenuation weight of the corresponding personnel feature file in the hot storage area to obtain the comprehensive matching score;

[0036] S34: Determine whether the overall matching score of the personnel feature file with the highest overall matching score in the hot storage area exceeds the preset re-identification confidence threshold;

[0037] S341: If the overall matching score exceeds the re-identification confidence threshold, the identity tracking identifier of the person's feature file will be determined as the identity backtracking matching result, and the search will be terminated;

[0038] S342: If the overall matching score does not exceed the re-identification confidence threshold, continue to execute the same retrieval process from step S32 to step S34 in the warm storage area;

[0039] S343: If there are still no personnel feature files in the warm storage area whose comprehensive matching score exceeds the re-identification confidence threshold, then continue to perform the same retrieval process from step S32 to step S34 in the cold storage area.

[0040] S344: If there are still no personnel feature files in the cold storage area whose comprehensive matching score exceeds the re-identification confidence threshold, then the identity backtracking matching result will be marked as a new identity tag, indicating that the person is a new person entering the scene for the first time;

[0041] S35: Output the identity backtracking matching result, which is an existing identity tracking identifier or a new identity marker.

[0042] S4: Based on the identity backtracking matching results, perform trajectory splicing and identity file backwriting and update for personnel who re-enter the full coverage area, merge the behavioral trajectories before and after leaving the venue into a complete continuous behavioral trajectory of the personnel, and simultaneously update the corresponding personnel feature files in the hierarchical persistent feature repository.

[0043] Further, step S4 includes:

[0044] S41: Determine the type of identity backtracking matching result;

[0045] S411: If the identity backtracking matching result is an existing identity tracking identifier, then extract the personnel feature file corresponding to the identity tracking identifier from the hierarchical persistent feature repository;

[0046] S412: If the identity backtracking matching result is a new identity tag, then assign a new identity tracking identifier to the person, create a new person feature profile and write it into the person feature profile set, and execute step S46;

[0047] S42: When the identity backtracking matching result is an existing identity tracking identifier, extract the last location coordinates and departure start timestamp recorded in the person's feature file before leaving the venue, and at the same time obtain the location coordinates and current timestamp of the person when they re-enter the full coverage area.

[0048] S43: Based on the last position coordinate before departure and the position coordinate upon re-entry, combined with the time interval between the departure start timestamp and the current timestamp, generate the interpolated trajectory segment during departure;

[0049] Specifically, the interpolation trajectory segment is generated as follows: taking the last position coordinate before departure as the starting point and the position coordinate upon re-entry as the ending point, and based on the constraints of the passable area in the scene spatial layout, an estimated position coordinate sequence during departure is generated using linear interpolation or passable path planning. A calculation mark is added to each coordinate point in the estimated position coordinate sequence to distinguish it from the actual observed position coordinate.

[0050] S44: Insert the interpolated trajectory segment between the historical location coordinate sequence before leaving the personnel's feature file and the real-time location coordinate after re-entry, forming a complete continuous behavior trajectory of the personnel.

[0051] S45: Perform a write-back update operation on the personnel feature profile: reset the departure attenuation weight to the initial value, move the personnel feature profile from the current storage level back to the personnel feature profile set as an active tracking target, and at the same time perform incremental weighted average update on the fused appearance feature vector in the personnel feature profile using the feature vector to be matched extracted in the current frame.

[0052] Specifically, the incremental weighted average update method is as follows: multiply the original fused appearance feature vector in the personnel feature file by the historical retention coefficient, add the feature vector to be matched extracted in the current frame multiplied by the current update coefficient, and write the sum of the two as the updated fused appearance feature vector back to the personnel feature file, wherein the sum of the historical retention coefficient and the current update coefficient is equal to one.

[0053] S46: The updated personnel feature profile is synchronously written into the personnel feature profile set. Subsequent frames continue to perform target detection and position update in step S1 on the personnel, so as to realize the continuous extension of the personnel's continuous behavior trajectory.

[0054] A continuous personnel tracking system for densely populated scenes, used to implement the aforementioned continuous personnel tracking method for densely populated scenes, the system comprising:

[0055] The acquisition and processing module is used to acquire multi-view image frame sequences within the coverage area of ​​multiple cameras, perform view coordinate mapping and global stitching on the multi-view image frame sequences to obtain a global stitched image frame sequence, and perform target detection and appearance feature extraction based on the global stitched image frame sequence to obtain a set of personnel feature profiles.

[0056] Hierarchical storage module: used to establish departure attenuation weights based on personnel feature profile sets, build a hierarchical persistent feature repository, and write the feature profiles of departing personnel into the hierarchical persistent feature repository according to the departure attenuation weights.

[0057] Re-identification module: used to extract the feature vector to be matched for newly entered personnel, and perform hierarchical cascaded re-identification retrieval in the hierarchical persistent feature repository according to priority to obtain the identity backtracking matching result;

[0058] The trajectory update module is used to perform trajectory splicing and identity file write-back updates based on the identity backtracking matching results, generate complete continuous behavior trajectories of personnel, and synchronously update the hierarchical persistent feature repository.

[0059] Compared to existing technologies, the advantages of this invention are as follows: This invention achieves simultaneous acquisition from multiple cameras via a network time protocol, ensuring temporal consistency of image frames from different perspectives at the same time. This provides a foundation for global stitching and coordinate unification. Combined with pixel-level registration and weighted fusion, it generates a global stitched image covering the entire scene, eliminating blind spots and improving the accuracy of personnel detection in dense scenes. This invention employs multi-granularity appearance feature extraction, fusing three complementary features: face, gait, and body shape. This ensures that effective features are available for identity matching under various occlusion and pose conditions in dense scenes, solving the problem of single feature extraction easily failing in dense scenes. This invention constructs a hierarchical persistent feature repository. Through a gradient storage strategy, it achieves a balance between retaining the identity information of departing personnel and controlling memory usage. Combined with hierarchical cascaded re-identification retrieval, it prioritizes searching high-confidence hot storage areas, effectively reducing the computational load of retrieval. Combined with departure attenuation weights, it avoids erroneous matching caused by expired features, solving the problem of identity reassignment and trajectory splitting upon return due to feature deletion after personnel leave. This invention generates physically reasonable interpolated trajectories based on scene passable area masks and A-satellite path planning, and incrementally weighted updates appearance features while taking into account feature stability and real-time performance, achieving complete connection of trajectories before and after departure, which can effectively improve the continuity and accuracy of visual personnel tracking in dense scenes. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a flowchart of the method for continuous tracking of people in dense scenes in this invention;

[0062] Figure 2 This is a schematic diagram of simultaneous data acquisition by multiple cameras in an embodiment of the present invention;

[0063] Figure 3 This is a schematic diagram illustrating the establishment of the scene global coordinate system in an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of weighted fusion splicing in an embodiment of the present invention;

[0065] Figure 5 This is a schematic diagram of multi-granularity appearance feature extraction in an embodiment of the present invention;

[0066] Figure 6 This is a schematic diagram of a hierarchical persistent feature repository in an embodiment of the present invention;

[0067] Figure 7 This is a functional block diagram of the continuous personnel tracking system in dense scenes in this invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example 1:

[0070] Please see Figure 1 As shown, this embodiment provides a method for continuous tracking of people in dense scenes, including:

[0071] S1: Collect multi-view image frame sequences within the coverage area of ​​multiple cameras, perform view coordinate mapping and global stitching on the multi-view image frame sequences to obtain a global stitched image frame sequence, and perform target detection and appearance feature extraction on all personnel in the current scene based on the global stitched image frame sequence to obtain a personnel feature profile set;

[0072] Further, step S1 includes:

[0073] S11: Obtain the sequence of multi-view image frames output by the matching multi-camera acquisition unit at the same time. Each image frame carries the intrinsic parameter matrix and extrinsic parameter matrix of the corresponding camera.

[0074] Further, step S11 includes:

[0075] S111: The accompanying multi-camera acquisition unit is deployed in the ceiling area of ​​the smart canteen or the top area of ​​the shelves in the smart warehouse. Each camera acquires image frames from the current viewpoint at a fixed frame rate. The fixed frame rate is determined based on the matching requirements between the movement speed of people in the scene and the image resolution. If the frame rate is too low, the displacement of people between adjacent frames will be too large, making it impossible to complete the inter-frame position matching. If the frame rate is too high, it will increase the processing load of the edge computing unit. For example, the fixed frame rate can be set to 25 frames per second. All cameras in the accompanying multi-camera acquisition unit are synchronized with the system clock of the edge computing unit through the network time protocol. Synchronous acquisition means that all cameras complete image sampling and record the timestamp at the same time based on the same clock signal, ensuring that image frames from different viewpoints are strictly aligned on the timeline. When the same person appears in the field of view of two cameras at the same time, the image frames acquired by the two cameras at the same time can capture the person's posture and position at the same instant.

[0076] S112: For each image frame in the multi-view image frame sequence, read the pre-calibrated intrinsic and extrinsic parameter matrices of the camera to which the image frame belongs. The intrinsic parameter matrix contains the camera's focal length, principal point coordinates, and distortion coefficients, describing the camera's internal optical imaging characteristics. The extrinsic parameter matrix contains the camera's rotation matrix and translation vector in the scene's global coordinate system, describing the camera's installation position and orientation in three-dimensional space. Both the intrinsic and extrinsic parameter matrices are obtained during the system deployment phase using the checkerboard calibration method and stored in the edge computing unit's configuration file.

[0077] S113: Each image frame, along with its corresponding intrinsic parameter matrix, extrinsic parameter matrix, and acquisition timestamp, is encapsulated into a parameterized image record. At the same time, the parameterized image records output by all cameras are arranged according to camera number to form a synchronous parameterized image record group. These synchronous parameterized image record groups from consecutive moments are arranged chronologically to form a multi-view image frame sequence. See also... Figure 2 This is a schematic diagram of simultaneous acquisition by multiple cameras provided in an embodiment of this application. Figure 2 As shown, the edge computing unit sends synchronization instructions to the supporting multi-camera acquisition unit deployed in the ceiling area of ​​the smart canteen or warehouse through the system clock and Network Time Protocol (NTP) to ensure that all cameras (camera one to camera N) are strictly aligned for image sampling at time nodes such as T1 and T2. Figure 2 This demonstrates a group of parameterized image records output at the same time T1, including current-view image frames from each camera, intrinsic and extrinsic parameter matrices, and the same timestamp. In dense acquisition scenarios, millisecond-level deviations in video streams from different viewpoints can lead to spatial misalignment and target duplication or missed detection during subsequent feature stitching. Figure 2 The revealed synchronous acquisition mechanism based on the unified distribution of the system clock ensures the absolute consistency of all images on the time axis from the physical acquisition source, laying a solid data foundation for spatiotemporal alignment for subsequent calculation processes such as coordinate mapping, blind spot elimination, and continuous feature tracking of personnel.

[0078] Specifically, the synchronous acquisition operation of multiple camera acquisition units in step S11 is the temporal basis for subsequent full-domain stitching and personnel tracking. In the smart canteen scenario, the flow of people is fast and the density is high during peak dining hours. If there is a millisecond-level deviation in the acquisition time of different cameras, the same person will present different postures and positions in image frames from different perspectives. This will cause spatial misalignment in the pixel-level registration in the subsequent step S13, and consequently, cause duplicate detection or missed detection in the overlapping areas of the target detection in step S14. By achieving synchronous acquisition of all cameras through the network time protocol, the temporal consistency of image frames from different perspectives at the same time is guaranteed, providing an accurate temporal alignment basis for subsequent coordinate mapping and full-domain stitching.

[0079] S12: Based on the intrinsic and extrinsic parameter matrices of each camera, calculate the overlapping area of ​​the viewpoints and the coordinate transformation relationship between adjacent cameras, and establish a unified global coordinate system for the scene;

[0080] Further, step S12 includes:

[0081] S121: Extract two parameterized image records corresponding to adjacent cameras from the synchronous parameterized image recording group. Calculate the relative rotation matrix and relative translation vector between the two cameras based on the rotation matrix and translation vector in their respective extrinsic parameter matrices. The relative rotation matrix is ​​obtained by multiplying the rotation matrix of the first camera by the transpose of the rotation matrix of the second camera. The relative translation vector is obtained by subtracting the product of the transpose of the rotation matrix of the first camera and the translation vector of the first camera from the translation vector of the second camera.

[0082] S122: Based on the intrinsic and extrinsic parameter matrices of the two cameras, project the field of view of the two cameras onto the ground plane of the scene, calculate the intersection of the two projected areas, and the intersection is the view overlap area. The view overlap area is represented by the set of polygon vertex coordinates in the global coordinate system of the scene.

[0083] S123: The coordinate system of camera number 1 in the multi-camera acquisition unit is selected as the reference origin of the scene global coordinate system. The remaining cameras map their local coordinates to the scene global coordinate system through the rotation matrix and translation vector in their respective extrinsic parameter matrices, thus establishing a unified scene global coordinate system. The scene global coordinate system uses the scene ground plane as the horizontal reference plane and the direction perpendicular to the ground as the height axis. See also Figure 3 This is a schematic diagram of the establishment of the scene global coordinate system provided in an embodiment of this application. For example... Figure 3 As shown, it intuitively depicts how the field-of-view projections of the first camera and the adjacent second camera overlap on the ground plane of the scene, and establishes a unified three-dimensional spatial reference. Figure 3 The document details the boundaries of each camera's field-of-view projections, the overlapping polygonal regions resulting from their intersections, and the set of vertex coordinates representing the physical extent of the overlap. Using the first camera as the reference origin (combining the horizontal and height axes), the document extracts the relative translation vectors and rotation matrices of adjacent cameras as extrinsic parameters and performs coordinate mapping transformations to merge the originally independent local coordinate systems into a unified global system. In wide-area, densely populated environments, inconsistent coordinate information often disrupts the continuous movement trajectories of personnel. This system establishment and coordinate transformation process eliminates spatial reference islands caused by the different installation positions of multiple cameras, ensuring that the calculated coordinates of personnel landing points in any frame can be accurately measured in the same three-dimensional space, further improving the reliability of Euclidean distance determination during inter-frame position matching.

[0084] Specifically, the scene global coordinate system established in step S12 is the foundation for achieving spatial unification of multi-view images. In the smart canteen scenario, multiple cameras are distributed at different positions on the ceiling, each with its own independent imaging coordinate system. Without establishing a unified scene global coordinate system, the same person detected by different cameras will have different position coordinates, making it impossible to determine whether the two coordinates point to the same person. Through the relative rotation matrix and relative translation vector calculated in step S121, and the unified scene global coordinate system established in step S123, subsequent step S13 can align image frames from different perspectives to the same spatial reference system for stitching, and step S14 can output the person's position coordinates in the unified coordinate system. Only then can the inter-frame position matching in step S2 calculate the Euclidean distance based on the same coordinate system. Without the coordinate unification in step S12, the position coordinates output in step S14 will be in different coordinate systems, the Euclidean distance calculation in step S21 will lose its physical meaning, and the inter-frame position matching will completely fail.

[0085] S13: In the scene global coordinate system, perform pixel-level registration and weighted fusion stitching on adjacent image frames with overlapping view areas in the multi-view image frame sequence to eliminate view blind spots and generate a global stitched image frame sequence covering the entire scene.

[0086] Further, step S13 includes:

[0087] S131: For each parameterized image record in the synchronous parameterized image record group, use the intrinsic parameter matrix in the parameterized image record to perform distortion correction on the image frame, eliminate the pixel position offset caused by lens radial distortion and tangential distortion, and obtain the corrected image frame.

[0088] S132: For two corrected image frames corresponding to adjacent cameras, using the relative rotation matrix and relative translation vector calculated in step S121, the pixel coordinates in the second corrected image frame are mapped to the coordinate plane where the first corrected image frame is located through homography transformation, thus completing pixel-level registration and obtaining a registered image frame pair.

[0089] S133: For the overlapping viewpoint regions in the registered image frame pairs, pixel-level fusion is performed using a distance-weighted fusion method. This method involves: for each pixel within the overlapping viewpoint region, calculating the distance from that pixel to the boundary of the first corrected image frame and the distance to the boundary of the second corrected image frame. Image frames farther from the boundary are assigned a higher fusion weight, while those closer are assigned a lower fusion weight. The pixel value of the pixel in each image frame is multiplied by its respective fusion weight and then summed to obtain the fused pixel value. Pixels outside the overlapping viewpoint region directly retain the pixel values ​​from the source image frames. The fused overlapping viewpoint region is then stitched together with the non-overlapping regions on both sides to generate a fully stitched image frame. See also... Figure 4 This is a schematic diagram of weighted fusion and splicing provided in an embodiment of this application. For example... Figure 4 As shown, Figure 4 This paper demonstrates the computational logic for pixel-level registration and transitional stitching in overlapping regions of viewpoints. The upper part depicts the first and second corrected image frames and their overlapping areas, indicating the "closer" and "farther" directions for extracting individual pixels from the two side boundaries. The middle part details the distance-weighted calculation rules through a fusion weighted line graph, showing the smooth transition trend of weights from 1.0 to 0. The lower part presents the final generated global stitched image frame, where original pixels are retained in non-overlapping areas, and a smooth transition gradient is formed in overlapping areas. In actual coverage areas, brightness differences and small residuals between adjacent shots often produce obvious illumination jumps and edge artifacts at overlapping boundaries, which can easily interfere with target detection algorithms and lead to misidentification as person bounding boxes. Through this graphical distance-weighted fusion mechanism, reliable pixels far from the boundary dominate the fusion result, greatly eliminating visual discontinuities caused by abrupt stitching, thereby significantly improving the quality of the global stitched image input to the target network.

[0090] Specifically, the reason for using distance-weighted fusion instead of direct stitching in step S13 is that in the overlapping view area, the images of the same physical area from the two cameras have brightness differences and slight registration residuals. Directly taking the pixel value from one of the cameras would cause obvious brightness jumps and edge artifacts at the stitching boundary. These artifacts would be misidentified as part of the person bounding box in the target detection in step S14, reducing detection accuracy. Distance-weighted fusion allows pixels far from the image boundary to dominate the fusion result, ensuring a smooth transition of pixel values ​​in the stitching transition area, eliminating brightness jumps, and guaranteeing the input image quality for subsequent target detection.

[0091] S14: Perform person target detection on each frame of the global stitched image frame sequence, and obtain the bounding box coordinates of each detected person and its position coordinates in the scene global coordinate system;

[0092] Specifically, the personnel target detection employs a single-stage target detection network based on an anchor-frame mechanism. This network takes a globally stitched image frame as input and outputs the bounding box coordinates and detection confidence scores for all personnel in the image. The bounding box coordinates are represented by the top-left x-coordinate, top-left y-coordinate, bounding box width, and bounding box height in a pixel coordinate system. The detection confidence score represents the probability that the bounding box contains a personnel target. Bounding boxes with detection confidence scores below a detection confidence threshold are filtered. This threshold is determined to minimize the number of false detection boxes while ensuring no real personnel targets are missed; for example, the detection confidence threshold can be set to 0.6. Non-maximum suppression is performed on the filtered bounding boxes to eliminate redundant bounding boxes generated by repeated detection of the same person. The midpoint pixel coordinates of the bottom edge of each retained bounding box are taken as the foot landing pixel coordinates of the person. These foot landing pixel coordinates are then converted to position coordinates in the scene's global coordinate system using the ground plane homography matrix in the scene's global coordinate system.

[0093] S15: Perform multi-granularity appearance feature extraction on the bounding box region of each inspected person to obtain a fused appearance feature vector;

[0094] Further, step S15 includes:

[0095] S151: Perform face feature encoding on the face region within the bounding box area to obtain a face feature vector. Specifically, firstly, a cascaded convolutional face detector is used to locate the face region within the bounding box area. The face region is represented by the coordinates of the upper left and lower right corners of the face bounding box. If no face region is detected within the bounding box area, the face feature vector is set to all zeros and marked as a missing face. If a face region is detected, the face region is scale-normalized and cropped to a fixed size before being input into a pre-trained face feature encoding network. The face feature encoding network takes the normalized cropped face image as input and outputs a 128-dimensional face feature vector. The face feature vector is then normalized using the L2 norm to make its magnitude 1.

[0096] S152: Perform gait cycle analysis and gait feature encoding on the whole-body region within the bounding box area to obtain a gait feature vector. Specifically, the gait cycle analysis requires continuous frame data. The bounding box regions of the same person within the current frame and the preceding preset window frames are arranged in chronological order to form a gait image sequence. The preset window frame number is determined based on the minimum number of frames required to cover a complete gait cycle; for example, the preset window frame number can be set to 30 frames. Human contour extraction is performed on each frame in the gait image sequence to obtain a binary contour sequence. This binary contour sequence is input into the gait energy map generation module. The gait energy map generation module averages the pixel positions of the binary contours of all frames in the binary contour sequence to generate a gait energy map. The gait energy map represents the average motion posture distribution of a person within a gait cycle. The gait energy map is input into a pre-trained gait feature encoding network, which outputs a 64-dimensional gait feature vector. This gait feature vector is normalized using the L2 norm to make its magnitude 1. If the number of consecutive frames preceding the current frame is less than the preset window frame number, the gait feature vector is set to an all-zero vector and a gait missing flag is marked.

[0097] S153: Perform clothing texture and body contour encoding on the torso region within the bounding box region to obtain a body feature vector. Specifically, within the bounding box region, based on the human keypoint detection results, the area from below the shoulders to above the knees is cropped as the torso region. The torso region is normalized and cropped to a fixed size before being input into a pre-trained body feature encoding network. The body feature encoding network takes the normalized cropped torso image as input, extracts clothing color histogram features, clothing texture features, and body contour features, and concatenates them to output a 64-dimensional body feature vector. The body feature vector is then normalized using the L2 norm to make its magnitude 1.

[0098] S154: The facial feature vector, gait feature vector, and body shape feature vector are concatenated in the order of facial feature vector first, gait feature vector second, and body shape feature vector last to generate a fused appearance feature vector for the person. The total dimensions of the fused appearance feature vector are 256 dimensions: 128 + 64 + 64. See also Figure 5 This is a schematic diagram of multi-granularity appearance feature extraction provided in an embodiment of this application. For example... Figure 5The diagram illustrates the hierarchical flow and dimensional concatenation process from the original target local region to the generation of a structured vector. The left side shows the spatial collaborative extraction relationship of the full-body region, face region, and torso region divided from the detected person's bounding box. The middle and right sections show the independent feature abstraction and dimensionality reduction of these three regions through face feature encoding (128-dimensional), body feature encoding (64-dimensional), and gait cycle analysis and gait feature encoding (64-dimensional), respectively. Finally, at the bottom, these are concatenated in a fixed dimensional order to form a 256-dimensional fused appearance feature vector. In crowded environments or in canteens and warehouses where people's postures frequently change, relying solely on easily occluded or non-frontal facial features for identification often leads to immediate tracking interruptions. The multi-granularity complementary architecture shown in this diagram fully utilizes the relatively stable torso color features and gait features with temporal periodic characteristics as effective fallback information, ensuring that even under adverse observation conditions such as people looking down, turning around, or being partially occluded, the system still has sufficient identification dimensions to accurately trace the person's identity.

[0099] Specifically, the reason for using multi-granularity appearance feature extraction instead of single facial or body feature extraction in step S15 is that in the dense crowd scene of a smart canteen, people frequently look down and turn around while eating, resulting in frequent situations where the face area is obscured or at a non-frontal angle. If only facial feature vectors are relied upon for identity matching, the feature matching ability will be completely lost in frames where the face is missing. Gait feature vectors are generated based on the temporal accumulation information of a person's walking posture, which does not depend on the face orientation and can still provide identity differentiation ability when the face is obscured. Body feature vectors are generated based on clothing texture and body contour, which can still provide auxiliary differentiation ability when the gait information is not significant when the person is standing still or waiting in line. The three granularities of features complement each other, ensuring that at least one effective feature can be used for identity matching under various postures and occlusion conditions in dense scenes. The fused appearance feature vector output in step S15 is directly bound to the personnel feature profile in step S16 and serves as the core basis for identity matching in the inter-frame position matching in step S2 and the hierarchical cascaded re-identification retrieval in step S3. If the multi-granularity appearance feature extraction in step S15 is missing, the inter-frame position matching in step S2 will only rely on the Euclidean distance of the position coordinates. When the position coordinates of multiple people are close in a dense crowd, it will be impossible to distinguish different individuals, resulting in the incorrect exchange of identity tracking tags.

[0100] S16: Assign a unique identity tracking identifier to each person being detected, and bind the identity tracking identifier with the corresponding fused appearance feature vector, the position coordinates of the current frame and the timestamp to form a person feature profile. All person feature profiles constitute a set of person feature profiles.

[0101] Specifically, the identity tracking identifier is represented by a 64-bit unsigned integer, generated by a globally incrementing counter maintained by the edge computing unit. Whenever a new person is detected who cannot be matched with an existing personnel feature profile, the global incrementing counter is incremented by 1, and the current count value is used as the person's identity tracking identifier. A personnel feature profile's data structure is a structured record containing the following fields: identity tracking identifier, fused appearance feature vector, historical location coordinate sequence, historical timestamp sequence, current state flag, and departure attenuation weight. The current state flag can be either active or absent, initially set to active. The historical location coordinate sequence is a list of the person's location coordinates in the scene's global coordinate system, stored chronologically. The historical timestamp sequence is a list of timestamps corresponding one-to-one with the historical location coordinate sequence. The departure attenuation weight is initially set to 1.0. All personnel feature profiles are stored in the memory of the edge computing unit, forming a personnel feature profile set.

[0102] S2: Based on the personnel feature profile set, establish an departure attenuation weight for each personnel feature profile, and construct a hierarchical persistent feature repository. When the tracked personnel leave the full coverage area, write the corresponding personnel feature profile into the hierarchical persistent feature repository according to the departure attenuation weight to obtain the hierarchical persistent feature repository.

[0103] Further, step S2 includes:

[0104] S21: Perform inter-frame position matching for each personnel feature file in the personnel feature file set between consecutive frames, and determine whether the personnel corresponding to each personnel feature file still appears in the latest frame of the current global stitched image frame sequence;

[0105] Further, step S21 includes:

[0106] S211: Obtain the position coordinates of all detected persons output in step S14 and the fused appearance feature vectors of all detected persons output in step S15 in the current frame. At the same time, read the latest position coordinates and fused appearance feature vectors of all personnel feature files marked as active in the personnel feature file set.

[0107] S212: Calculate the Euclidean distance between the position coordinates of each detected person in the current frame and the latest position coordinate of each active person feature file in the personnel feature file set. Simultaneously, calculate the cosine similarity between the fused appearance feature vector of each detected person in the current frame and the fused appearance feature vector in each active person feature file. Pairs with an Euclidean distance lower than the position matching threshold and a cosine similarity higher than the appearance matching threshold are identified as the same person. The position matching threshold is determined based on the maximum reasonable displacement distance between people in adjacent frames. This maximum reasonable displacement distance is determined by the product of the person's maximum walking speed and the inter-frame time interval. For example, when the person's maximum walking speed is 2 meters per second and the inter-frame time interval is 0.04 seconds, the position matching threshold can be set to 0.08 meters. The appearance matching threshold is determined based on the minimum stable similarity of the appearance features of the same person in adjacent frames. For example, the appearance matching threshold can be set to 0.75.

[0108] S213: When multiple individuals simultaneously meet the matching criteria for the same personnel feature profile, the Hungarian algorithm is used to find the globally optimal matching solution. The Hungarian algorithm uses a weighted combination of Euclidean distance and cosine similarity as the cost matrix element. This weighted combination is Euclidean distance multiplied by a distance cost coefficient minus cosine similarity multiplied by an appearance cost coefficient, outputting the one-to-one matching result with the minimum total cost. For successfully matched personnel feature profiles, the location coordinates and timestamp of the corresponding individual in the current frame are appended to the historical location coordinate sequence and historical timestamp sequence of that personnel feature profile.

[0109] Specifically, step S21 uses both the Euclidean distance of position coordinates and the cosine similarity of the fused appearance feature vectors for inter-frame position matching, rather than using only one metric. This is because in densely populated scenarios like smart canteens, multiple people queuing at food windows are very close together. Relying solely on the Euclidean distance of position coordinates can easily lead to confusion between adjacent people's position coordinates within the position matching threshold, resulting in incorrect swapping of identity tracking markers. Introducing the cosine similarity of the fused appearance feature vectors as a second constraint allows for differentiation between multiple people in close proximity based on their appearance differences. Furthermore, step S213 introduces the Hungarian algorithm to solve for the globally optimal match because many-to-many matching conflicts may occur in dense scenarios. A greedy, one-to-one matching strategy cannot guarantee global optimality and may cause locally optimal matches to crowd out the correct matching opportunities for other people. The Hungarian algorithm ensures that the total cost of all matching pairs is minimized, reducing matching errors to the greatest extent possible.

[0110] S22: If a personnel feature profile is not matched with any detected personnel in the current frame within a consecutive preset number of frames, then the personnel feature profile is marked as leaving the field and the departure start timestamp is recorded.

[0111] Specifically, the determination of the consecutive preset frame number is based on excluding situations where a person is temporarily not detected due to brief obstruction but is still actually in the scene. In a smart canteen scenario, a person being briefly obstructed by someone queuing in front usually lasts no more than 0.5 seconds. For example, the consecutive preset frame number can be set to 15 frames. When a person feature profile currently marked as active fails to be successfully matched with any detected person in the most recent 15 consecutive frames by the inter-frame position matching in step S21, the current status of the person feature profile is changed from active to absent. At the same time, the timestamp of the last successful match is recorded as the departure start timestamp and written into the person feature profile.

[0112] S23: Calculate the departure attenuation weight for the personnel feature files marked as departing;

[0113] Specifically, the calculation method for the exit decay weight is as follows: taking the exit start timestamp as the starting point, reading the current system timestamp, calculating the exit duration as the difference between the current system timestamp and the exit start timestamp, and substituting the exit duration as the independent variable into the exponential decay function to calculate the exit decay weight. The expression for the exponential decay function is:

[0114]

[0115] in, As the exit decay weight, For the duration of the absence, This is the attenuation rate parameter. The attenuation rate parameter is determined by the scene type parameter. The attenuation rate parameter for the cafeteria scene is smaller than that for the warehouse scene. This means that in the cafeteria scene, the probability of people returning after leaving the camera's coverage area is higher, and the return time interval is longer. Therefore, the attenuation weight decreases more slowly to retain identity information for a longer period. In the warehouse scene, people's work routes are relatively fixed, and they usually return within a short time or do not return for a long time after leaving the coverage area. Therefore, the attenuation rate parameter is set larger to release storage resources occupied by people who will not return more quickly. For example, the attenuation rate parameter for the cafeteria scene can be set to 0.005, and the attenuation rate parameter for the warehouse scene can be set to 0.01. For example, assuming a person's absence in the cafeteria scene lasts for 60 seconds, the attenuation weight is calculated as follows: This indicates that the matching confidence level of the personnel's profile is 74.1% of the initial value; if the absence duration is 300 seconds, the absence attenuation weight is calculated as follows: The match reliability dropped to 22.3%.

[0116] S24: Construct a hierarchical persistent feature repository, which includes three levels: hot storage area, warm storage area and cold storage area;

[0117] Further, step S24 includes:

[0118] S241: When a person's feature profile is marked as being out of the scene and the duration of the absence does not exceed a first duration threshold, the person's feature profile is migrated from the person's feature profile set to the hot storage area. The person's feature profile in the hot storage area retains a complete fused appearance feature vector and all historical location coordinate sequences. The first duration threshold is determined based on the upper limit of the time when a person is most likely to return after briefly leaving the coverage area within the scene. For example, the first duration threshold for a cafeteria scene can be set to 120 seconds, and the first duration threshold for a warehouse scene can be set to 60 seconds.

[0119] S242: When the duration of absence exceeds a first duration threshold but not a second duration threshold, the personnel feature file is transferred from the hot storage area to the warm storage area. The personnel feature file in the warm storage area retains the fused appearance feature vector and a recently preset number of historical location coordinates. The recently preset number is determined by retaining sufficient trajectory information to support the generation of interpolated trajectory segments in step S4. For example, the recently preset number can be set to 10 historical location coordinates. The warm storage area deletes earlier historical location coordinates exceeding the recently preset number to free up storage space. The second duration threshold is determined by the upper limit of the time a person might return after a moderate period of absence within the scene. For example, the second duration threshold for a cafeteria scene can be set to 600 seconds, and the second duration threshold for a warehouse scene can be set to 300 seconds.

[0120] S243: When the duration of absence exceeds the second duration threshold but not the third duration threshold, the personnel feature file is transferred from the warm storage area to the cold storage area. The personnel feature file in the cold storage area retains only the fused appearance feature vector and identity tracking identifier, deleting all historical location coordinate sequences and historical timestamp sequences. The third duration threshold is determined based on the maximum duration of a single operational cycle in the scenario. For example, the third duration threshold for a canteen scenario can be set to 3600 seconds corresponding to a dining period, and the third duration threshold for a warehouse scenario can be set to 1800 seconds.

[0121] S244: When the duration of absence exceeds the third duration threshold, delete the personnel profile from the cold storage area, release all storage space occupied by the personnel profile, and reclaim the identity tracking identifier to mark it as a historically used identifier. See also Figure 6 This is a schematic diagram of the hierarchical persistent feature repository provided in an embodiment of this application. For example... Figure 6 As shown, this figure visualizes the evolution of the storage state of departing personnel's characteristic files across different time periods, constrained by gradient decay. (Transverse) Figure 6The horizontal dashed lines represent the first, second, and third duration thresholds. As the vertical axis indicates an increasing duration of absence and an exponentially decreasing weighting of absence decay, the storage priority of tracking files is progressively downgraded: initially, a "hot storage area" retaining all historical continuous tracks and complete features; then a "warm storage area" with a moderately reduced number of tracks; next, a "cold storage area" containing only identity identifiers and feature information; and finally, a complete release of resources and recycling of used identity identifiers after exceeding the bottom limit due to timeout. In existing schemes, the simple destruction strategy that treats both long and short durations indiscriminately often segments individuals who temporarily leave a specific coverage area and return within a short period into multiple new identities. Figure 6 The revealed gradient aging and feature hierarchical persistence mechanism, while maintaining the lowest possible minimum resident memory consumption of computing units, greatly improves the tracking accuracy when personnel repeatedly move back and forth within the monitoring field of view, ensuring the long-span temporal sequence of the overall flow trajectory chain.

[0122] Specifically, the core purpose of constructing the hierarchical persistent feature repository in step S2 is to solve the core technical problem addressed by this patent: when a person temporarily leaves the coverage area of ​​the full-area camera and then re-enters the coverage area, the consistency of the original identity binding cannot be maintained. Existing tracking algorithms, in order to control memory usage, directly delete the feature-identity mapping relationship after a person leaves, resulting in the same person being assigned a new identity tracking identifier upon re-entry, and the complete behavioral trajectory being split into multiple independent trajectories. The hierarchical persistent feature repository achieves a balance between retaining the identity information of departing personnel and controlling system memory usage through a gradient storage strategy of three levels: hot storage area, warm storage area, and cold storage area. The hot storage area retains complete information to cope with the situation of rapid return after a short absence; the warm storage area retains sufficient identity matching information while moderately simplifying historical trajectory data; and the cold storage area retains only the minimum identity features to cope with possible return after a long absence. The mechanism of decreasing departure attenuation weight over time avoids incorrect matching of expired feature information of people who have been away for a long time in the re-identification retrieval step S3. This is because as the departure time increases, the appearance of people may change, leading to a decrease in feature similarity. The introduction of attenuation weight can reduce the matching weight of expired features. At the same time, the differentiated time threshold configuration for the canteen and warehouse scenarios reflects cross-scenario adaptation optimization. In the canteen scenario, people may leave the food pick-up area to go to the dining area and then return to get more food. The departure time is longer, but the probability of returning is higher. Therefore, all three time thresholds are greater than those for the warehouse scenario. The dependency relationship between step S2 and step S1 is as follows: the inter-frame position matching in step S2 depends on the position coordinates output in step S14 and the fused appearance feature vector output in step S15. Without the global stitching and multi-granularity feature extraction in step S1, step S2 will not be able to obtain effective matching input data. The output of step S2, namely the hierarchical persistent feature repository, is directly used by the hierarchical cascaded re-identification retrieval in step S3. If the hierarchical persistent storage mechanism of step S2 is missing, step S3 will not be able to find the feature files of departing personnel at any storage level, and the re-identification retrieval will be completely ineffective.

[0123] S3: When a person enters the full coverage area, the fused appearance feature vector of the person is extracted as the feature vector to be matched. In the hierarchical persistent feature repository, the feature vector to be matched is subjected to hierarchical cascaded re-identification retrieval according to the priority order of hot storage area, warm storage area and cold storage area to obtain the identity backtracking matching result.

[0124] Further, step S3 includes:

[0125] S31: Perform the same multi-granularity appearance feature extraction as in step S15 on newly entered personnel in the full coverage area to generate the personnel's feature vector to be matched;

[0126] Specifically, the method for determining newly entered personnel in the full coverage area is as follows: After inter-frame position matching in step S21, the personnel detected in the current frame in step S14 are determined to be newly entered personnel in the full coverage area if they do not match any active personnel feature files in the personnel feature file set. The bounding box region of this personnel is then sequentially processed by step S151 (face feature encoding), step S152 (gait cycle analysis and gait feature encoding), step S153 (clothing texture and body contour encoding), and step S154 (dimensional concatenation) to generate a 256-dimensional feature vector to be matched.

[0127] S32: In the hot storage area of ​​the hierarchical persistent feature repository, calculate the weighted similarity between the feature vector to be matched and the fused appearance feature vector of each personnel feature profile in the hot storage area;

[0128] Further, step S32 includes:

[0129] S321: The feature vector to be matched is split into a face component, a gait component, and a body shape component according to the dimensional concatenation order in step S154. The face component consists of dimensions 1 to 128 of the feature vector to be matched, the gait component consists of dimensions 129 to 192 of the feature vector to be matched, and the body shape component consists of dimensions 193 to 256 of the feature vector to be matched. The same splitting operation is performed on the fused appearance feature vector of each personnel feature file in the hot storage area.

[0130] S322: Calculate the cosine similarity between the face component in the feature vector to be matched and the face component in the personnel feature file, the cosine similarity between the gait component in the feature vector to be matched and the gait component in the personnel feature file, and the cosine similarity between the body shape component in the feature vector to be matched and the body shape component in the personnel feature file.

[0131] S323: Check whether there are missing face or gait markers in the feature vector to be matched and the personnel feature file. If the face component in either the feature vector to be matched or the personnel feature file is marked as a missing face marker, the face weight coefficient is set to 0 and the remaining weights are redistributed according to the original ratio of the gait weight coefficient to the body shape weight coefficient. If the gait component is marked as a missing gait marker, the gait weight coefficient is set to 0 and the remaining weights are redistributed according to the original ratio of the face weight coefficient to the body shape weight coefficient. In the absence of missing markers, the face weight coefficient, gait weight coefficient, and body shape weight coefficient are determined based on the distinguishing ability and stability of the three features in dense scenes. Face features have the strongest individual distinguishing ability, so the face weight coefficient is the largest. Gait features have a stable periodic pattern in the walking state, so the gait weight coefficient is the second largest. Body shape features are most affected by lighting and viewing angle, so the body shape weight coefficient is the smallest. For example, the face weight coefficient can be set to 0.5, the gait weight coefficient can be set to 0.3, and the body shape weight coefficient can be set to 0.2. The weighted similarity is obtained by multiplying the cosine similarity of the face component by the face weight coefficient, the cosine similarity of the gait component by the gait weight coefficient, and the cosine similarity of the body shape component by the body shape weight coefficient, and then summing them.

[0132] For example, suppose the cosine similarity between the face component in the feature vector to be matched and the face component in a certain personnel feature file is 0.92, the cosine similarity between the gait component and the body shape component is 0.85, and there is no missing marker, then the weighted similarity is calculated as 0.92 multiplied by 0.5 plus 0.85 multiplied by 0.3 plus 0.78 multiplied by 0.2 equals 0.46 plus 0.255 plus 0.156 equals 0.871.

[0133] S33: Multiply the weighted similarity by the departure attenuation weight of the corresponding personnel feature file in the hot storage area to obtain the comprehensive matching score;

[0134] Specifically, the physical meaning of the comprehensive matching score is as follows: the weighted similarity reflects the degree of matching between the currently entering and leaving personnel in terms of appearance features, and the departure attenuation weight reflects the timeliness and reliability of the departing personnel's profile. The comprehensive matching score, after multiplying the two, considers both the degree of feature matching and the timeliness and reliability. Even if the weighted similarity is high, the comprehensive matching score will decrease due to the decrease in the departure attenuation weight for personnel profiles with longer departure times, thus avoiding incorrect matching due to changes in personnel appearance after a long period of absence. For example, continuing the numerical example above, if the departure duration of the personnel profile is 60 seconds and the scene is a cafeteria, then the departure attenuation weight is 0.741, and the comprehensive matching score is 0.871 multiplied by 0.741, which equals 0.645.

[0135] S34: Determine whether the overall matching score of the personnel feature file with the highest overall matching score in the hot storage area exceeds the preset re-identification confidence threshold;

[0136] Further, step S34 includes:

[0137] S341: If the overall matching score exceeds the re-identification confidence threshold, the identity tracking identifier of the person's feature file is determined as the identity backtracking matching result, and the search is terminated. The re-identification confidence threshold is determined based on maximizing the matching recall rate while ensuring the matching accuracy. If the re-identification confidence threshold is too high, matches that are actually for the same person will be rejected and a new identity tracking identifier will be assigned. If the re-identification confidence threshold is too low, different people will be incorrectly matched. For example, the re-identification confidence threshold can be set to 0.55.

[0138] S342: If the overall matching score does not exceed the re-identification confidence threshold, then continue to perform the weighted similarity calculation of steps S321 to S323 and the overall matching score calculation of step S33 on the fused appearance feature vector of the feature vector to be matched and each personnel feature file in the warm storage area. Determine whether the overall matching score of the personnel feature file with the highest overall matching score in the warm storage area exceeds the re-identification confidence threshold. If it does, then determine the identity tracking identifier of the personnel feature file as the identity backtracking matching result and terminate the search.

[0139] S343: If there are still no personnel feature files in the warm storage area with a comprehensive matching score exceeding the re-identification confidence threshold, then continue to perform the same search process in the cold storage area. If the comprehensive matching score of the personnel feature file with the highest comprehensive matching score in the cold storage area exceeds the re-identification confidence threshold, then the identity tracking identifier of that personnel feature file is determined as the identity backtracking matching result and the search is terminated. If there are still no personnel feature files in the cold storage area with a comprehensive matching score exceeding the re-identification confidence threshold, then the identity backtracking matching result is marked as a new identity identifier, indicating that the person is a new person entering the scene for the first time.

[0140] Specifically, in step S3, a hierarchical cascaded re-identification retrieval is performed according to the priority order of the hot storage area, warm storage area, and cold storage area, rather than mixing all personnel feature files from all three levels for a unified retrieval. This is because the hot storage area stores complete feature files of people who have briefly left the area; these people have the highest probability of returning to the scene, and the timeliness and reliability of their feature files are the strongest. Prioritizing retrieval in the hot storage area allows for the identification matching of most returning personnel with minimal computational overhead. The retrieval is only extended to the warm and cold storage areas when no matching is found in the hot storage area. This avoids traversing all personnel feature files across all three levels every time a new person enters, significantly reducing the computational load during peak hours when the number of people leaving the smart canteen is large. The weight redistribution mechanism for missing face and gait markers in step S323 ensures that matching can still be performed using remaining available features when some features are unavailable, avoiding the loss of correct matches due to an abnormally low overall weighted similarity caused by the absence of a certain granularity feature. The design in step S33, which multiplies the weighted similarity score by the departure attenuation weight, ensures that the re-identification retrieval considers not only the matching degree of appearance features but also the timeliness of the feature profile. Individuals who have been absent for an extended period will not be easily matched, even if their feature similarity is high. This mechanism avoids cross-personal mismatches caused by appearance similarity between different individuals. The collaborative relationship between steps S3 and S2 is as follows: Step S2 stores the feature profiles and departure attenuation weights of departing individuals through a hierarchical persistent feature repository. Step S3 then performs hierarchical cascaded re-identification retrieval based on this. If only the retrieval mechanism of step S3 is present without the hierarchical persistent storage of step S2, the feature profiles of departing individuals will be directly deleted after departure, and there will be no feature profiles available for matching. If only the storage mechanism of step S2 is present without the hierarchical cascaded retrieval of step S3, even if the feature profiles of departing individuals are saved, newly entering individuals cannot be identified as returning individuals and will still be assigned a new identity tracking identifier. The collaboration of these two steps achieves the persistent storage of departing individuals' identity information and the accurate identity backtracking of returning individuals.

[0141] S4: Based on the identity backtracking matching results, perform trajectory splicing and identity file backwriting and update for personnel who re-enter the full coverage area, merge the behavioral trajectories before and after leaving the venue into a complete continuous behavioral trajectory of the personnel, and simultaneously update the corresponding personnel feature files in the hierarchical persistent feature repository.

[0142] Further, step S4 includes:

[0143] S41: Determine the type of identity backtracking matching result;

[0144] Specifically, step S41 reads the identity backtracking matching result output in step S35 and determines whether the value of the identity backtracking matching result is an existing identity tracking identifier or a new identity tag.

[0145] S411: If the identity backtracking matching result is an existing identity tracking identifier, locate the storage level of the identity tracking identifier in the hierarchical persistent feature repository and extract the personnel feature file corresponding to the identity tracking identifier, and execute step S42.

[0146] S412: If the identity backtracking matching result is a new identity tag, then call the global incrementing counter in step S16 to assign a new identity tracking tag to the person, create a new person feature file, use the feature vector to be matched generated in step S31 as the fused appearance feature vector, use the position coordinates and timestamp of the current frame as the first record of the historical position coordinate sequence and historical timestamp sequence, set the current state tag to active state, set the departure attenuation weight to 1.0, write it into the person feature file set, and execute step S46.

[0147] S42: When the identity backtracking matching result is an existing identity tracking identifier, extract the last location coordinates and departure start timestamp recorded in the person's feature file before leaving the venue, and at the same time obtain the location coordinates and current timestamp of the person when they re-enter the full coverage area.

[0148] Further, step S42 includes:

[0149] S421: Read the last location coordinate from the extracted personnel feature file's historical location coordinate sequence as the last location coordinate before departure, and read the departure start timestamp. If the personnel feature file is located in the warm storage area, the historical location coordinate sequence retains the most recent preset number of historical location coordinates, and the one with the latest timestamp is taken as the last location coordinate before departure. If the personnel feature file is located in the cold storage area, the historical location coordinate sequence has been deleted, and the last location coordinate before departure is marked as a location missing marker.

[0150] S422: Obtain the position coordinates of the person in the current frame output by step S14 as the position coordinates when re-entering, and obtain the timestamp of the current frame as the current timestamp.

[0151] S43: Based on the last position coordinate before departure and the position coordinate upon re-entry, combined with the time interval between the departure start timestamp and the current timestamp, generate the interpolated trajectory segment during departure;

[0152] Further, step S43 includes:

[0153] S431: Calculate the departure time interval as the difference between the current timestamp and the departure start timestamp. Determine whether the last position coordinate before departure was marked as a missing position marker. If it was marked as a missing position marker, skip the generation of the interpolated trajectory segment and directly use the position coordinate at the time of re-entry as the starting point for continuing the continuous behavior trajectory of the person, and execute step S44.

[0154] S432: If the last location coordinate before leaving the scene is not marked as a missing location marker, then read the scene passable area mask pre-stored in the edge computing unit. The scene passable area mask is a binary grid map corresponding to the scene global coordinate system. The area with a grid value of 1 represents the area that people can pass through, and the area with a grid value of 0 represents the area occupied by obstacles. The obstacles include tables and food pick-up stations in the canteen scene and shelves in the warehouse scene.

[0155] S433: Starting from the last position coordinate before departure and ending at the position coordinate upon re-entry, the A* path planning algorithm is used on the scene's passable area mask to search for the shortest passable path from the starting point to the ending point. The A* path planning algorithm uses the Euclidean distance between grid nodes as a heuristic function and only expands the search among passable area nodes with a grid value of 1, outputting a sequence of shortest passable path coordinates composed of ordered path node coordinates. The shortest passable path coordinate sequence is sampled at equal time intervals according to the departure time interval. Each sampled coordinate point in the shortest passable path coordinate sequence is marked with a calculation marker to distinguish it from the actual observed position coordinates, resulting in an interpolated trajectory segment. The calculation marker is a Boolean field; a true value indicates that the position coordinates are calculated rather than actually observed.

[0156] Specifically, in step S43, the A* path planning algorithm is used in conjunction with a mask of the scene's passable areas to generate interpolated trajectory segments, rather than simply performing linear interpolation between the last position coordinates before departure and the position coordinates upon re-entry. This is because, in a smart canteen scenario, a person's position before departure might be near the food pick-up window, and their position upon re-entry might be near the restaurant entrance. The straight path between these two positions might cross the table area, which is impossible for a person to actually walk across. If a trajectory generated using linear interpolation crosses impassable areas, the trajectory is physically unreasonable and could lead the subsequent behavior analysis system to misjudge that the person is lingering in impassable areas. The A* path planning algorithm searches for the shortest passable path around obstacles based on the scene's passable area mask, generating interpolated trajectory segments that match the person's actual possible walking routes, ensuring the physical rationality of the trajectory data. Simultaneously, a calculation marker is added to each coordinate point of the interpolated trajectory segment, allowing the subsequent behavior analysis system to distinguish between the actual observed trajectory and the calculated trajectory. During detailed behavior analysis, the calculated trajectory segment can be ignored, preventing the calculated data from interfering with the accuracy of actual behavior judgment.

[0157] S44: Insert the interpolated trajectory segment between the historical location coordinate sequence before leaving the personnel's feature file and the real-time location coordinate after re-entry, forming a complete continuous behavior trajectory of the personnel.

[0158] Specifically, if step S43 generates an interpolated trajectory segment, all sampled coordinate points and their calculated markers in the interpolated trajectory segment are appended in chronological order to the last position coordinate before departure in the historical position coordinate sequence of the personnel feature file. Then, the position coordinates upon re-entry and the current timestamp are appended to the interpolated trajectory segment, so that the historical position coordinate sequence of the personnel feature file forms a continuous trajectory of personnel behavior that is temporally and spatially coherent from the first entry into the scene to the current moment. If step S431 skips the generation of the interpolated trajectory segment, the position coordinates upon re-entry and the current timestamp are directly appended to the end of the historical position coordinate sequence of the personnel feature file. There is a temporal and spatial discontinuity between the two trajectories, which is reflected in the historical position coordinate sequence as the time gap between the departure start timestamp and the current timestamp.

[0159] S45: Perform a write-back update operation on the personnel's profile;

[0160] Further, step S45 includes:

[0161] S451: Reset the departure attenuation weight of the personnel feature file to the initial value of 1.0, and change the current status flag of the personnel feature file from departure status to active status.

[0162] S452: Move the personnel feature profile from its current storage level back to the personnel feature profile set as an active tracking target. The move-back operation is as follows: delete the copy of the personnel feature profile from the corresponding level of the hierarchical persistent feature repository, and write the updated personnel feature profile into the personnel feature profile set.

[0163] S453: The fused appearance feature vector in the personnel feature file is updated using the feature vector to be matched extracted in the current frame with an incremental weighted average. The incremental weighted average update is calculated as follows: the original fused appearance feature vector in the personnel feature file is multiplied by the historical retention coefficient, and the feature vector to be matched extracted in the current frame is multiplied by the current update coefficient. The sum of the two is written back to the personnel feature file as the updated fused appearance feature vector. The sum of the historical retention coefficient and the current update coefficient is equal to 1. The historical retention coefficient is determined based on absorbing the latest appearance changes while retaining the long-term stable appearance features of the personnel. For example, the historical retention coefficient can be set to 0.8, and the current update coefficient can be set to 0.2. The physical meaning of the incremental weighted average update is: the appearance of personnel may change locally before and after leaving the scene due to reasons such as changing coats or wearing masks. If the original fused appearance feature vector is used completely, the matching may fail in the inter-frame position matching of subsequent frames due to the large difference in appearance. If the feature vector to be matched in the current frame is used to replace the original fused appearance feature vector, the stable appearance information accumulated by the personnel over a long period of time will be lost. By using a weighted average with a historical retention coefficient of 0.8, the updated fused appearance feature vector retains historically stable features with 80% weight and incorporates the latest current features with 20% weight, achieving a balance between feature stability and real-time performance.

[0164] For example, suppose the value of a certain dimension of the original fused appearance feature vector in a certain personnel feature file is 0.65, and the value of the corresponding dimension of the feature vector to be matched extracted in the current frame is 0.72. Then, after the update, the value of that dimension is 0.65 multiplied by 0.8 plus 0.72 multiplied by 0.2 equals 0.52 plus 0.144 equals 0.664.

[0165] S46: The updated personnel feature profile is synchronously written into the personnel feature profile set. Subsequent frames continue to perform target detection and position update in step S1 on the personnel, so as to realize the continuous extension of the personnel's continuous behavior trajectory.

[0166] Specifically, after step S46 is executed, the person exists in an active state in the personnel feature file set. When the person is detected in step S14 in each subsequent full-domain stitched image frame, step S21 will perform inter-frame position matching based on the updated fused appearance feature vector and the latest position coordinates. After successful matching, the new position coordinates and timestamp will be appended to the historical position coordinate sequence and historical timestamp sequence of the person's feature file. The person's continuous behavior trajectory continues to extend over time until the person leaves the field again, triggering the departure status mark in step S22.

[0167] Specifically, step S4 achieves a complete connection between the behavioral trajectories before and after departure through trajectory stitching and identity file back-writing updates, solving the core problem of the same person's behavioral data being split into multiple independent trajectories. In step S43, the interpolated trajectory segments generated based on the scene's passable area mask and the A-Star path planning algorithm ensure the physical rationality of the trajectory estimation during departure. The incremental weighted average update in step S453 ensures that the fused appearance feature vector maintains effective matching capability even after local changes in the person's appearance. The dependency between steps S4 and S3 is as follows: the identity backtracking matching result of step S4 is entirely derived from the hierarchical cascaded re-identification retrieval output of step S3. If the identity backtracking matching result output by step S3 is an existing identity tracking identifier, step S4 can perform the trajectory stitching operation; if step S3 outputs a new identity marker because no match is found in the hierarchical persistent feature repository, step S4 creates a new person feature file. The circular dependency between step S4 and step S1 is as follows: After step S4 completes the write-back update, the number of active tracking targets and their feature content in the personnel feature file set change. These changes directly affect the matching candidate set when performing inter-frame position matching in step S21 of subsequent frames, forming a continuous cyclical tracking process from step S1 to step S4. The four steps from S1 to S4 form a closed-loop collaboration: step S1 provides the basic data for full-domain image acquisition and personnel feature extraction; step S2 establishes a hierarchical and persistent storage mechanism for the features of departing personnel; step S3 performs identity backtracking matching when personnel return; and step S4 transforms the matching results into specific operations for trajectory stitching and file updating. The absence of any of these four steps will lead to a break in the continuous tracking chain: without step S1, no features can be extracted; without step S2, information on departing personnel is lost; without step S3, returning personnel cannot be identified; and without step S4, the matching results cannot be transformed into trajectory stitching. This four-step closed loop supports continuous identity tracking of personnel throughout their entire journey in a smart canteen, from entering the canteen, queuing for food, leaving the food collection area, eating in the dining area, and then returning to the food collection area for a second meal. In a smart warehouse, it supports continuous identity maintenance of personnel as they move between shelves, temporarily leave the monitored area to go to an uncovered rest area, and then return to the work area. This provides a complete and continuous foundation of personnel behavior data for subsequent behavior analysis, safety control, and operational decisions.

[0168] For example, a smart canteen scenario illustrates the complete workflow. A military canteen deploys six cameras during lunchtime to cover the food pickup area and main passageways. The camera frame rate is set to 25 frames per second, and the edge computing unit simultaneously receives the six image frames. At 11:30 AM, personnel A enters the food pickup area for the first time. Step S1 performs coordinate mapping and global stitching on the six image frames to generate a global stitched image frame. Step S14 detects the bounding box of personnel A in the global stitched image frame. Step S15 extracts personnel A's facial feature vector, gait feature vector, and body shape feature vector and concatenates them into a 256-dimensional fused appearance feature vector. Step S16 assigns personnel A an identity tracking identifier of 10086, creates a personnel feature profile, and writes it into the personnel feature profile set.

[0169] At 11:35, Person A finished picking up their food and left the food pick-up area to go to the dining area, which was located in a blind spot of the camera coverage. Step S21 failed to match Person A within 15 consecutive frames. Step S22 marked Person A's personnel feature profile as "offline" and recorded the departure start timestamp as 11:35:00. Step S23 continuously calculated Person A's departure attenuation weight; after 30 seconds, the departure attenuation weight was... Since the duration of absence is 30 seconds, which does not exceed the first duration threshold of 120 seconds in the cafeteria scenario, step S241 writes the personnel characteristic file of personnel A into the hot storage area.

[0170] At 11:42, Person A returned from the dining area to the food pick-up area to refill their meal. Step S14 detected a new person entering the current frame. Step S21's inter-frame location matching did not find a match in the active target set of personnel feature profiles. Step S31 extracted the feature vector to be matched for this person. Step S32 calculated the weighted similarity between the feature vector to be matched and Person A's personnel feature profile in the hot storage area, finding it to be 0.871. Step S33 multiplied the weighted similarity by the departure decay weight (departure time 420 seconds). However, since Person A had been transferred from the hot storage area to the warm storage area in step S242 more than 120 seconds after leaving the site, step S34 did not find a match in the hot storage area. Step S342 retrieved Person A's personnel feature file in the warm storage area and calculated the comprehensive matching score as 0.871 multiplied by The score is 0.106, which does not exceed the re-identification confidence threshold of 0.55. It's important to note that the lower decay weight in the example above is due to the 420-second absence time. If Person A returns in a shorter time, the overall matching score will be higher. Assuming that in a real-world scenario Person A's weighted similarity is 0.93 and returns after 180 seconds, the overall matching score would be 0.93 multiplied by... The value is 0.379, still lower than 0.55. Considering the long departure time of people in the cafeteria scenario, the decay rate parameter can be further reduced to 0.002. Therefore, the decay weight after 180 seconds is... The overall matching score is 0.93 multiplied by 0.698, which equals 0.649, exceeding the re-identification confidence threshold of 0.55. Step S341 identifies Person A's identity tracking identifier 10086 as the identity backtracking matching result. Step S43 uses Person A's last location coordinates before leaving (coordinates in front of the food pick-up window) as the starting point and the location coordinates upon re-entry (coordinates at the entrance of the food pick-up area) as the ending point, generating an interpolated trajectory segment bypassing the dining area using the A* path planning algorithm on the scene's passable area mask. Step S44 inserts the interpolated trajectory segment into Person A's historical location coordinate sequence. Step S45 resets the departure attenuation weight to 1.0, migrates Person A's personnel feature file back to the personnel feature file set, performs incremental weighted average updates, and continues tracking Person A's food ordering behavior, achieving continuous identity tracking of Person A from their first entry into the cafeteria to the completion of their food ordering.

[0171] Example 2:

[0172] This embodiment, based on Embodiment 1, provides a continuous personnel tracking system for densely populated scenes, such as... Figure 7 As shown, it includes:

[0173] The acquisition and processing module is used to acquire multi-view image frame sequences within the coverage area of ​​multiple cameras, perform view coordinate mapping and global stitching on the multi-view image frame sequences to obtain a global stitched image frame sequence, and perform target detection and appearance feature extraction based on the global stitched image frame sequence to obtain a set of personnel feature profiles.

[0174] Hierarchical storage module: used to establish departure attenuation weights based on the personnel feature profile set, construct a hierarchical persistent feature repository, and write the feature profiles of departing personnel into the hierarchical persistent feature repository according to the departure attenuation weights;

[0175] Re-identification module: used to extract the feature vector to be matched for newly entered personnel, and perform hierarchical cascaded re-identification retrieval in the hierarchical persistent feature repository according to priority to obtain the identity backtracking matching result;

[0176] Trajectory Update Module: Used to perform trajectory splicing and identity file write-back update based on the identity backtracking matching results, generate complete continuous behavior trajectory of personnel, and synchronously update the hierarchical persistent feature repository.

Claims

1. A method for continuous tracking of people in dense scenes, characterized in that, The method includes: S1: Collect multi-view image frame sequences within the coverage area of ​​multiple cameras, perform view coordinate mapping and global stitching on the multi-view image frame sequences to obtain a global stitched image frame sequence, and perform target detection and appearance feature extraction on all personnel in the current scene based on the global stitched image frame sequence to obtain a personnel feature profile set; S2: Based on the personnel feature profile set, establish an departure attenuation weight for each personnel feature profile, construct a hierarchical persistent feature repository, and when the tracked personnel leave the full coverage area, write the corresponding personnel feature profile into the hierarchical persistent feature repository according to the departure attenuation weight to obtain the hierarchical persistent feature repository. S3: When a person enters the full coverage area, the fused appearance feature vector of the person is extracted as the feature vector to be matched. In the hierarchical persistent feature repository, the feature vector to be matched is subjected to hierarchical cascaded re-identification retrieval in the priority order of hot storage area, warm storage area and cold storage area to obtain the identity backtracking matching result. S4: Based on the identity backtracking and matching results, perform trajectory splicing and identity file backwriting and update for personnel who re-enter the full coverage area, merge the behavioral trajectories before and after leaving the venue into a complete continuous behavioral trajectory of the personnel, and simultaneously update the corresponding personnel feature files in the hierarchical persistent feature repository.

2. The method for continuous tracking of people in dense scenes according to claim 1, characterized in that, S1 includes: The system acquires a sequence of multi-view image frames output simultaneously by a multi-camera acquisition unit. Each image frame carries the intrinsic and extrinsic parameter matrices of the corresponding camera.

3. The method for continuous tracking of people in densely populated scenes according to claim 2, characterized in that, The supporting multi-camera acquisition unit is deployed in the ceiling area of ​​the smart canteen or the top area of ​​the shelves in the smart warehouse. All cameras are synchronized with the system clock of the edge computing unit through the Network Time Protocol.

4. The method for continuous tracking of people in dense scenes according to claim 1, characterized in that, S1 further includes performing multi-granularity appearance feature extraction on the bounding box region of each detected person to obtain a fused appearance feature vector. The multi-granularity appearance feature extraction includes: Perform facial feature encoding on the face region within the bounding box to obtain a facial feature vector; Gait cycle analysis and gait feature encoding are performed on the bounding box regions of consecutive frames to obtain gait feature vectors; Clothing texture and body shape contour encoding are performed on the torso region within the bounding box to obtain the body feature vector.

5. The method for continuous tracking of people in dense scenes according to claim 4, characterized in that, The fused appearance feature vector is obtained by sequentially concatenating the face feature vector, gait feature vector, and body shape feature vector, with a total dimension of 256.

6. The method for continuous tracking of people in dense scenes according to claim 1, characterized in that, The departure attenuation weight is calculated based on the departure duration using an exponential decay function, and the attenuation rate parameter is preset according to the scenario type.

7. The method for continuous tracking of people in dense scenes according to claim 1, characterized in that, The hierarchical persistent feature repository includes three storage levels: hot storage area, warm storage area, and cold storage area.

8. The method for continuous tracking of people in dense scenes according to claim 7, characterized in that, The hierarchical migration rules for the hierarchical persistent feature repository are as follows: Items marked as being out of the field and whose out-of-field duration does not exceed the first duration threshold are stored in the hot storage area. Items exceeding the first duration threshold but not the second duration threshold are transferred to the warm storage area; Items exceeding the second duration threshold but not the third duration threshold are transferred to the cold storage area; The profiles of the individuals whose profiles exceed the third time threshold will be deleted.

9. The method for continuous tracking of people in dense scenes according to claim 1, characterized in that, In the hierarchical cascaded re-identification retrieval, after calculating the weighted similarity between the feature to be matched and the stored file, the weighted similarity is multiplied by the departure attenuation weight of the corresponding personnel feature file to obtain the comprehensive matching score.

10. The method for continuous tracking of people in dense scenes according to claim 1, characterized in that, The steps for generating the interpolated trajectory segment during the departure period include: Starting from the last position coordinate before leaving the field, and ending from the position coordinate when re-entering the field; The shortest passable path is searched using the A* path planning algorithm on the mask of the passable area in the scene. The shortest passable path is sampled at equal time intervals to obtain the interpolated trajectory segment.

11. The method for continuous tracking of people in dense scenes according to claim 1, characterized in that, When writing back and updating personnel feature profiles, an incremental weighted average update is performed on the fused appearance feature vector, retaining the original historical stable features while incorporating the latest current features.

12. A continuous personnel tracking system for dense scenes, used to implement the continuous personnel tracking method for dense scenes according to any one of claims 1-11, characterized in that, The system includes: The acquisition and processing module is used to acquire multi-view image frame sequences within the coverage area of ​​multiple cameras, perform view coordinate mapping and global stitching on the multi-view image frame sequences to obtain a global stitched image frame sequence, and perform target detection and appearance feature extraction based on the global stitched image frame sequence to obtain a set of personnel feature profiles. Hierarchical storage module: used to establish departure attenuation weights based on the personnel feature profile set, construct a hierarchical persistent feature repository, and write the feature profiles of departing personnel into the hierarchical persistent feature repository according to the departure attenuation weights; Re-identification module: used to extract the feature vector to be matched for newly entered personnel, and perform hierarchical cascaded re-identification retrieval in the hierarchical persistent feature repository according to priority to obtain the identity backtracking matching result; Trajectory Update Module: Used to perform trajectory splicing and identity file write-back update based on the identity backtracking matching results, generate complete continuous behavior trajectory of personnel, and synchronously update the hierarchical persistent feature repository.

Images