Cross-scene historical trajectory tracing method based on character feature uniqueness

Abstract

The invention belongs to the technical field of artificial intelligence and computer vision, and particularly discloses a cross-scene historical track tracing method based on character feature uniqueness, which comprises the following steps: extracting static features and dynamic features of a character, and fusing to generate a unique feature code; performing cross-scene feature matching among the cameras of the plurality of scenes; according to a cross-scene feature matching result, four-tuple nodes of different scenes are connected in series according to a time sequence, the four-tuple nodes are associated to form a historical track of the same person, and association conditions include that the feature similarity is not lower than 95%, the time difference is not larger than 2.4 seconds, and the spatial distance is not larger than 0.5 m; and in response to a retrieval condition input by a user, retrieving a matched track from the historical tracks, and outputting a visual historical track, the visual historical track comprising the track marked with the key time point and the associated video clip. According to the invention, scene limitation is broken through, and efficient and accurate cross-scene historical track tracing can be realized.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and computer vision technology, and more specifically, relates to a method for tracing historical trajectories across scenes based on the uniqueness of human characteristics. Background Technology

[0002] With the widespread deployment of surveillance equipment, streets, shopping malls, and parks are now covered by numerous cameras. Existing video processing technologies typically employ person trajectory tracking schemes based on single features such as facial features or clothing color. However, this type of technology has significant limitations: most schemes can only record trajectories within the coverage area of ​​a single camera, failing to effectively connect across cameras or areas, resulting in isolated trajectory tracks. Furthermore, the features relied upon are susceptible to changes in lighting, occlusion, and posture, leading to poor stability and causing person matching failures. When it is necessary to query the historical movement trajectory of a specific person, existing methods require manual review of video recordings from different cameras segment by segment, which is inefficient and makes it difficult to accurately correlate the same person across different scenes.

[0003] Therefore, how to overcome scenario limitations and achieve efficient and accurate cross-scenario historical trajectory tracing is an urgent problem to be solved. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this application is to provide a method for tracing historical trajectories across scenarios based on the uniqueness of personal characteristics, which breaks through the limitations of scenarios and can achieve efficient and accurate tracing of historical trajectories across scenarios.

[0005] To achieve the above objectives, firstly, this application provides a method for tracing historical trajectories across scenarios based on the uniqueness of personal characteristics, comprising the following steps: S10, extract static and dynamic features of a person from video streams from multiple cameras, and fuse the static and dynamic features to generate a unique feature code for the person; S20, based on the unique feature encoding, perform cross-scene feature matching among cameras in multiple scenes. The cross-scene feature matching includes using KD tree for coarse matching to obtain candidate feature encoding and using triplet loss for fine matching to determine the person matching. S30, the trajectory of the person under each camera is divided into a quadruple node consisting of a timestamp, geographic location coordinates, unique feature code and camera identifier, and according to the cross-scene feature matching result, the quadruple nodes of different scenes are linked in chronological order to be associated with the historical trajectory of the same person. The association conditions include feature similarity of not less than 95%, time difference of not more than 2.4 seconds and spatial distance of not more than 0.5 meters. S40, in response to the search conditions input by the user, the search conditions include feature samples, time range and scene range, retrieve matching trajectories from the historical trajectories, and output visualized historical trajectories, the visualized historical trajectories include trajectories marked with key time points and associated video clips.

[0006] The historical trajectory tracing method based on the uniqueness of human characteristics provided in this application has the following effects: By fusing static and dynamic features to generate a unique feature code in step S10, the problem of poor stability caused by the reliance on a single feature in traditional technologies can be solved; Step S20 adopts a strategy combining coarse and fine matching, and introduces scene calibration and time window prediction, which can improve the accuracy and efficiency of cross-scene matching; Step S30 constructs a trajectory based on strict association conditions and integrates deduplication and correction mechanisms to ensure trajectory continuity; Step S40 supports multi-condition retrieval and visualization output, which can realize the transformation from manual investigation to AI-level second-level tracing. The synergistic effect of each step breaks through the limitations of a single scene. By utilizing feature uniqueness and matching optimization, efficient and accurate cross-scene historical trajectory tracing can be achieved.

[0007] As a further preferred embodiment, in step S10, the static features are extracted using the Dino model, and the static features include facial detail texture and the height ratio and body contour features of the person; the dynamic features are extracted using 3D CNN and Transformer models, and the dynamic features include the person's gait rhythm, arm swing frequency and walking posture behavior features.

[0008] As a further preferred option, in step S20, before the cross-scene feature matching step, spatial coordinate calibration and illumination calibration are performed on multiple cameras to obtain the positional relationship of the cameras and the overlapping area of ​​their fields of view.

[0009] As a further preferred option, a trajectory deduplication step is also included, which uses the DBSCAN algorithm to deduplicate the repeated trajectory nodes of the same person in the overlapping area of ​​multiple cameras. The spatial clustering radius of the DBSCAN algorithm is 0.5 meters and the minimum number of samples is 3.

[0010] As a further preferred option, an abnormal trajectory correction step is also included, in which linear interpolation is used to complete the abnormal segment when the spatial distance between adjacent trajectory nodes exceeds 2 meters or the time difference is less than 0.1 seconds.

[0011] Secondly, this application provides a cross-scene historical trajectory tracing system based on the uniqueness of personal characteristics, including: The feature extraction module is used to extract static and dynamic features of a person from video streams from multiple cameras, and fuse the static and dynamic features to generate a unique feature code of the person. The feature matching module is used to perform cross-scene feature matching among cameras in multiple scenes based on the unique feature encoding. The cross-scene feature matching includes using KD trees for coarse matching to obtain candidate feature encodings and using triplet loss for fine matching to determine the person matching. The trajectory construction module is used to decompose the trajectory of a person under each camera into a quadruple node consisting of a timestamp, geographic location coordinates, unique feature code, and camera identifier. Based on the cross-scene feature matching results, the quadruple nodes of different scenes are linked together in chronological order to associate them as the historical trajectory of the same person. The association conditions include feature similarity of not less than 95%, time difference of not more than 2.4 seconds, and spatial distance of not more than 0.5 meters. The historical tracing module is used to respond to user-input search conditions, including feature samples, time range, and scene range, retrieve matching trajectories from the historical trajectories, and output visualized historical trajectories, which include trajectories marked with key time points and associated video clips.

[0012] As a further preferred embodiment, in the historical tracing module, the associated video segments are loaded in chunks based on timestamps, with each 10-second segment loaded. The video encoding adopts the H.265 format, and the loading time of the first frame does not exceed 0.8 seconds.

[0013] As a further preferred embodiment, the historical tracing module supports multi-condition combined retrieval based on feature samples, time range, and scene range.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the cross-scene historical trajectory tracing method based on the uniqueness of human characteristics as described in any of the above-mentioned methods.

[0015] Fourthly, this application provides a computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, implement the cross-scene historical trajectory tracing method based on the uniqueness of human characteristics as described above.

[0016] It is understandable that the beneficial effects of the second, third and fourth aspects mentioned above can be found in the relevant descriptions in the first aspect above, and will not be repeated here. Attached Figure Description

[0017] Figure 1 This is a flowchart of the cross-scene historical trajectory tracing method based on the uniqueness of human characteristics provided in this application; Figure 2 This is a diagram illustrating the overall technical architecture provided in the embodiments of this application; Figure 3This is a flowchart of the human feature extraction process provided in the embodiments of this application; Figure 4 This is a cross-scene trajectory visualization interface provided in the embodiments of this application, which displays trajectory nodes and video clips. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be understood that, in the description of this application, the terms "first" and "second," etc., are used to distinguish different objects, rather than to describe a specific order of objects.

[0020] like Figure 1 As shown, this application provides a method for tracing historical trajectories across scenarios based on the uniqueness of personal characteristics. This method can be widely applied to scenarios such as security monitoring, intelligent transportation, and commercial passenger flow analysis. It includes steps S10 to S40, which are detailed below: Step S10: Extract static and dynamic features of the person from the video streams of multiple cameras, and fuse the static and dynamic features to generate a unique feature code for the person.

[0021] In step S10, static features can be extracted using the Dino model, specifically including biometric features such as facial detail texture and morphological features such as height proportion and body contour. Dynamic features can be extracted using a 3D CNN combined with a Transformer architecture, including behavioral features such as gait rhythm, arm swing frequency, and walking posture. An attention mechanism is used to assign weights to static and dynamic features, ultimately generating a unique 128-dimensional feature code. This ensures that the feature similarity of the same person in different scenes is no less than 95%, and the similarity between different people is no more than 10%, thereby improving feature stability and uniqueness and overcoming the feature matching failure problem caused by factors such as lighting and occlusion in traditional technologies.

[0022] Step S20: Based on unique feature encoding, cross-scene feature matching is performed between cameras in multiple scenes. Cross-scene feature matching includes using KD-tree for coarse matching to obtain candidate feature encoding, and using triplet loss for fine matching to determine the person matching.

[0023] In step S20, spatial coordinate calibration and illumination calibration can be performed on multiple cameras to obtain the positional relationship of the cameras and the overlapping area of ​​their fields of view, thereby reducing environmental interference.

[0024] In the coarse matching stage, the top 10 candidate feature codes with cosine similarity can be retrieved using a KD-tree, with a screening time of no more than 0.1 seconds per attempt. In the fine matching stage, triplet loss can be used to optimize feature distance calculation, ensuring that the distance between the feature to be matched and the positive sample feature is less than 0.15 and the distance with the negative sample feature is greater than 0.8, thereby improving feature discriminative power. Combined with the prediction of the time window based on the person's movement speed and timestamp, the matching time is no more than 0.5 seconds per attempt, thus achieving efficient and accurate cross-scene person association.

[0025] Step S30: The trajectory of a person under each camera is split into a quadruple node consisting of a timestamp, geographic coordinates, unique feature code, and camera identifier. Based on the cross-scene feature matching results, the quadruple nodes of different scenes are linked together in chronological order to form the historical trajectory of the same person. The association conditions include feature similarity of not less than 95%, time difference of not more than 2.4 seconds, and spatial distance of not more than 0.5 meters.

[0026] In step S30, the DBSCAN algorithm can be used to deduplicate repeated trajectory nodes of the same person in the overlapping area of ​​multiple cameras. The spatial clustering radius is 0.5 meters, the minimum number of samples is 3, and the duplicate node rate after deduplication does not exceed 0.5%, effectively reducing storage redundancy. Abnormal trajectory correction can also be performed. When the spatial distance between adjacent trajectory nodes exceeds 2 meters or the time difference is less than 0.1 seconds, linear interpolation is used to complete the abnormal segment, with a correction rate of not less than 98%, improving trajectory continuity and accuracy, and reducing the trajectory breakage rate from the traditional 12% to below 2%.

[0027] Step S40: In response to the user's input search criteria, which include feature samples, time range, and scene range, retrieve matching trajectories from historical trajectories and output visualized historical trajectories. The visualized historical trajectories include trajectories marked with key time points and associated video clips.

[0028] In step S40, historical tracing supports multi-condition retrieval based on feature samples, time range, and scene range, achieving a response time within seconds. The visualization output can overlay historical trajectories and annotate key time points. Video clips are encoded using H.265 and loaded in segments based on timestamps, with each 10-second segment loaded. The first frame loading time is no more than 0.8 seconds, reducing bandwidth usage and improving user experience.

[0029] The historical trajectory tracing method based on the uniqueness of human characteristics provided in this application has the following effects: By fusing static and dynamic features to generate a unique feature code in step S10, the problem of poor stability caused by the reliance on a single feature in traditional technologies can be solved; Step S20 adopts a strategy combining coarse and fine matching, and introduces scene calibration and time window prediction, which can improve the accuracy and efficiency of cross-scene matching; Step S30 constructs a trajectory based on strict association conditions and integrates deduplication and correction mechanisms to ensure trajectory continuity; Step S40 supports multi-condition retrieval and visualization output, which can realize the transformation from manual investigation to AI-level second-level tracing. The synergistic effect of each step breaks through the limitations of a single scene. By utilizing feature uniqueness and matching optimization, efficient and accurate cross-scene historical trajectory tracing can be achieved.

[0030] Based on the same inventive concept, this application also provides a cross-scene historical trajectory tracing system based on the uniqueness of human characteristics, including a feature extraction module, a feature matching module, a trajectory construction module, and a historical tracing module.

[0031] The feature extraction module is used to extract static and dynamic features of a person from video streams from multiple cameras, and to fuse the static and dynamic features to generate a unique feature code for the person.

[0032] In this application, the feature extraction module can specifically use the Dino model to extract static features, including facial details and body contours, and use 3D CNN and Transformer models to extract dynamic features, such as gait rhythm and arm swing frequency. The module uses an attention mechanism to fuse and generate a 128-dimensional unique feature code to ensure feature uniqueness and stability.

[0033] The feature matching module is used to perform cross-scene feature matching between cameras in multiple scenes based on unique feature encoding. Cross-scene feature matching includes using KD trees for coarse matching to obtain candidate feature encodings and using triplet loss for fine matching to determine the person matching.

[0034] In this application, the feature matching module can first perform scene calibration preprocessing, including spatial coordinate calibration and illumination calibration, to obtain the camera position relationship and the overlapping area of ​​the field of view; coarse matching can quickly retrieve candidate features through KD tree, fine matching uses triplet loss to optimize feature distance, and combines time window prediction to improve matching efficiency and accuracy.

[0035] The trajectory construction module is used to break down the trajectory of a person under each camera into a quadruple node consisting of a timestamp, geographic coordinates, unique feature code, and camera identifier. Based on the cross-scene feature matching results, the quadruple nodes of different scenes are linked together in chronological order to associate them as the historical trajectory of the same person. The association conditions include feature similarity of not less than 95%, time difference of not more than 2.4 seconds, and spatial distance of not more than 0.5 meters. In this application, the trajectory construction module can also integrate trajectory deduplication function, using the DBSCAN algorithm to cluster and deduplicate duplicate nodes in overlapping areas, as well as abnormal trajectory correction function, using linear interpolation to complete abnormal segments, ensuring trajectory continuity and integrity.

[0036] The historical tracing module is used to respond to user-input search conditions, including feature samples, time range, and scene range. It retrieves matching trajectories from historical trajectories and outputs visualized historical trajectories, which include trajectories marked with key time points and associated video clips.

[0037] In this application, the historical tracing module supports multi-condition combined retrieval, video segments are loaded in segments based on timestamps, each segment is loaded in 10-second chunks, the video encoding adopts H.265 format, and the first frame loading time does not exceed 0.8 seconds, achieving efficient retrieval and low-bandwidth visualization output.

[0038] It should be noted that the functions of each module provided in this application can be found in the detailed description of the methods provided above, and will not be repeated here.

[0039] In one embodiment, the technical solution to achieve the above objective can be as follows: In order to solve the problems of broken cross-scene character trajectory, unstable feature matching, and inefficient historical trajectory tracing, this embodiment provides a cross-scene historical trajectory tracing system based on the uniqueness of character features. The system extracts unique character features through AI algorithms, constructs a cross-scene trajectory association model, and achieves a breakthrough from "single-scene tracking" to "full-link tracing".

[0040] Specifically, such as Figure 2 As shown, the historical trajectory tracing system provided in this application works collaboratively through four major modules: "Dino feature extraction - cross-scene matching - trajectory construction - historical tracing". The core architecture is as follows: (1) Character Unique Feature Extraction Module like Figure 3 As shown, a fusion of "static features + dynamic features" is used to ensure feature uniqueness and stability. Static feature extraction: Based on the Dino model, extract biological features such as facial details and textures, as well as the height ratio and body contour features of the person; Dynamic feature extraction: Based on a 3D CNN + Transformer combined architecture, behavioral features such as gait rhythm, arm swing frequency, and walking posture of a person are extracted. Dynamic feature vectors are constructed through temporal video frame analysis. Feature fusion and encoding: Attention mechanism is used to assign weights to static and dynamic features, and finally a unique 128-dimensional feature code is generated to ensure that the feature similarity of the same person in different scenes is ≥95% and the similarity of different people is ≤10%.

[0041] (2) Cross-scene feature matching module An adaptive matching algorithm is designed to address the differences in scene conditions such as viewpoint, distance, and lighting among multiple cameras: Scene calibration preprocessing: By calibrating the spatial coordinates of multiple cameras, the positional relationship and overlapping area of ​​the field of view of different cameras are obtained; the illumination intensity of different scenes is unified to reduce environmental interference; Feature matching algorithm: Employs a two-stage strategy of "coarse matching + fine matching". Coarse matching: Input the Dino feature code of the person to be matched, retrieve the candidate feature codes of the top 10 cosine similarity in the KD tree, and the filtering time is ≤0.1 seconds / time, reducing the computational load of subsequent fine matching; Fine Matching: Optimize Dino Feature Distance Calculation - Using "feature to be matched, positive sample feature (same person), negative sample feature (similar person)" as a triple, the distance between the feature to be matched and the positive sample feature is <0.15, and the distance between the feature to be matched and the negative sample feature is >0.8, thus improving feature discrimination; eliminate "similar feature interference" from different people with similar clothing and body shape, and finally determine the unique matching object; Cross-scene connection logic: When a person enters scene B from scene A, based on the spatial overlap area of ​​the camera calibration, combined with the person's movement speed and timestamp, the time window for the person to enter scene B is predicted, narrowing the matching range and improving matching efficiency. The matching time is ≤0.5 seconds / time.

[0042] (3) Cross-scene trajectory construction and optimization module Trajectory node generation: The trajectory of a person under each camera is split into a four-tuple node of "time-location-feature-camera ID", for example, [2025-08-26 10:00:00, coordinates (120.5, 350.2, 10.0), feature code X, scene A camera 1];

[0043] Trajectory association and completion: Based on feature matching results, nodes from different scenes are linked together in chronological order. When the similarity between the Dino feature code of a new node and a historical node is ≥95%, and the spatial and temporal conditions meet the association criteria, the new node is automatically associated with the trajectory of the same person, forming a complete trajectory. The trajectory nodes are stored in a Pgsql database, which supports fast querying by time range and camera ID. The storage size of a single node is ≤1KB, and the data storage of 100 cameras for one month is approximately 80GB.

[0044] Trajectory deduplication and correction: Through clustering algorithms, duplicate trajectory nodes of the same person in overlapping areas of multiple cameras are eliminated; abnormal trajectories are corrected by combining scene geographic information.

[0045]

[0046] (4) Historical trajectory tracing module High-efficiency search interface: Supports uploading a frame of a person image and searching based on multiple conditions such as "feature sample", "time range", and "scene range"; Trajectory visualization output: Overlays the traced historical trajectory, annotates key time points (e.g., "10:05 Entering the construction area", "10:20 Leaving the construction area"), and generates corresponding trajectory video clips, such as... Figure 4 As shown.

[0047]

[0048] The following is a specific implementation example of this application: (a) Hardware environment

[0049] (II) Software Deployment Process Data acquisition layer: Accesses real-time video streams from cameras in various scenes via the RTSP protocol and stores the raw video in the format of "scene-camera ID-time"; Feature extraction layer: Deploy the fine-tuned Dino model at the edge nodes to extract static features of the person, extract both static and dynamic features of the person, and generate unique feature codes; Matching and Trajectory Layer: Deploy algorithms such as KD-tree retrieval to achieve cross-scene feature matching and trajectory construction, and store trajectory nodes in a PgSql database; Application layer: Develop a web-based visualization platform that provides historical trajectory retrieval and trajectory display functions, and supports permission management for administrators and ordinary users.

[0050] (III) Case Verification Taking the cross-scenario of "park-construction area-rest area" as an example: 1) When a person enters the frame from camera A in the construction site, the system extracts their feature code X and records the trajectory node [10:00, Park A (100,200), X]; 2) At 10:05, a person enters the construction area of ​​the park via camera B. The system confirms that the code X matches the person in park A through scene calibration and feature matching, and records the node [10:05, shopping mall B (50,150), X], linking the two trajectories. 3) At 10:20, the person leaves the construction area of ​​the park and enters the park rest area camera C. The system predicts that the entry time window is 10:20-10:22. The code X is quickly matched and the node [10:21, Park C (80,300), X] is recorded. 4) When a user uploads an image of a person from camera A in the park at 10:00 and searches for the trajectory from "10:00 to 11:00", the system returns the complete trajectory within 3 seconds and displays the path "Park A → Construction Area B → Rest Area C" on the map, supporting viewing the video of each node.

[0051] (iv) Supplementary Experimental Data 1. Quantitative parameters for experimental scenarios, taking the cross-scenario of "park-construction area-rest area" as an example.

[0052] 2. Comparison of basic indicators for ordinary scenes with normal lighting and occlusion rate ≤30%

[0053] 3. Verification of anti-interference capability in complex scenes with strong light / weak light / high occlusion

[0054] 4. Matching accuracy of similar appearance groups

[0055] The beneficial effects of this embodiment are: (1) Cross-scene trajectory continuity: Breaking through the limitation of "camera island", it achieves seamless connection of character trajectory in multiple scenes, with a trajectory integrity rate of ≥98%; Taking the Dino model as the core, it solves the pain points of "unstable features and cross-scene break" in traditional technology, improving the cross-scene feature matching accuracy by more than 33 percentage points and the trajectory integrity rate by more than 40 percentage points; (2) Feature uniqueness and stability: By integrating static and dynamic features, the ability to resist interference from lighting and occlusion is improved, and the accuracy of character matching is ≥96%; (3) High efficiency of historical tracing: The process has been upgraded from "manually checking video segment by segment" to "AI second-level retrieval". The historical trajectory tracing has been reduced from "hour-level" to "second-level". The data retrieval of 30 days from 100 cameras only takes 3 seconds. There is no need for manual checking of each camera. The tracing efficiency has been improved by more than 100 times and the labor cost has been reduced by 90%. (4) Multi-scenario applicability: It maintains high stability in complex scenarios and can be flexibly deployed in security, transportation, commerce and other fields, such as quickly locking the trajectory of suspects in security, analyzing customer shopping paths in commerce, and monitoring personnel activities in park management, adapting to the needs of different industries. (5) Data value: The generated complete trajectory data can support multi-dimensional analysis, such as customer movement heatmaps in commerce and suspect behavior patterns in security, providing data support for decision-making.

[0056] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for tracing historical trajectories across scenes based on the uniqueness of individual characteristics, characterized in that, Includes the following steps: S10, extract static and dynamic features of a person from video streams from multiple cameras, and fuse the static and dynamic features to generate a unique feature code for the person; S20, based on the unique feature encoding, perform cross-scene feature matching among cameras in multiple scenes. The cross-scene feature matching includes using KD tree for coarse matching to obtain candidate feature encoding and using triplet loss for fine matching to determine the person matching. S30, the trajectory of the person under each camera is divided into a quadruple node consisting of a timestamp, geographic location coordinates, unique feature code and camera identifier, and according to the cross-scene feature matching result, the quadruple nodes of different scenes are linked in chronological order to be associated with the historical trajectory of the same person. The association conditions include feature similarity of not less than 95%, time difference of not more than 2.4 seconds and spatial distance of not more than 0.5 meters. S40, in response to the search conditions input by the user, the search conditions include feature samples, time range and scene range, retrieve matching trajectories from the historical trajectories, and output visualized historical trajectories, the visualized historical trajectories include trajectories marked with key time points and associated video clips.

2. The historical trajectory tracing method based on the uniqueness of personal characteristics across scenes as described in claim 1, characterized in that, In step S10, the static features are extracted using the Dino model, and the static features include facial detail texture and the height ratio and body contour features of the person; the dynamic features are extracted using 3D CNN and Transformer models, and the dynamic features include the person's gait rhythm, arm swing frequency and walking posture behavior features.

3. The historical trajectory tracing method based on the uniqueness of personal characteristics across scenes as described in claim 1, characterized in that, In step S20, before the cross-scene feature matching step, spatial coordinate calibration and illumination calibration are performed on multiple cameras to obtain the positional relationship of the cameras and the overlapping area of ​​their fields of view.

4. The historical trajectory tracing method across scenes based on the uniqueness of personal characteristics as described in claim 1, characterized in that, It also includes a trajectory deduplication step, which uses the DBSCAN algorithm to deduplicate repeated trajectory nodes of the same person in the overlapping area of ​​multiple cameras. The spatial clustering radius of the DBSCAN algorithm is 0.5 meters and the minimum number of samples is 3.

5. The method for tracing historical trajectories across scenes based on the uniqueness of personal characteristics as described in claim 1, characterized in that, It also includes an abnormal trajectory correction step, in which linear interpolation is used to complete the abnormal segment when the spatial distance between adjacent trajectory nodes exceeds 2 meters or the time difference is less than 0.1 seconds.

6. A cross-scene historical trajectory tracing system based on the uniqueness of personal characteristics, characterized in that, include: The feature extraction module is used to extract static and dynamic features of a person from video streams from multiple cameras, and fuse the static and dynamic features to generate a unique feature code of the person. The feature matching module is used to perform cross-scene feature matching among cameras in multiple scenes based on the unique feature encoding. The cross-scene feature matching includes using KD trees for coarse matching to obtain candidate feature encodings and using triplet loss for fine matching to determine the person matching. The trajectory construction module is used to decompose the trajectory of a person under each camera into a quadruple node consisting of a timestamp, geographic location coordinates, unique feature code, and camera identifier. Based on the cross-scene feature matching results, the quadruple nodes of different scenes are linked together in chronological order to associate them as the historical trajectory of the same person. The association conditions include feature similarity of not less than 95%, time difference of not more than 2.4 seconds, and spatial distance of not more than 0.5 meters. The historical tracing module is used to respond to user-input search conditions, including feature samples, time range, and scene range, retrieve matching trajectories from the historical trajectories, and output visualized historical trajectories, which include trajectories marked with key time points and associated video clips.

7. The historical trajectory tracing system based on the uniqueness of personal characteristics across scenes as described in claim 6, characterized in that, In the historical tracing module, the associated video segments are loaded in chunks based on timestamps, with each 10-second segment loaded. The video encoding uses the H.265 format, and the first frame loading time does not exceed 0.8 seconds.

8. The historical trajectory tracing system based on the uniqueness of personal characteristics across scenes as described in claim 6, characterized in that, The historical tracing module supports multi-condition combined retrieval based on feature samples, time range, and scene range.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the cross-scene historical trajectory tracing method based on the uniqueness of human characteristics as described in any one of claims 1 to 5.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instruction is executed by the processor, it implements the historical trajectory tracing method based on the uniqueness of human characteristics across scenes as described in any one of claims 1 to 5.

Images