Pedestrian re-identification method based on path tracking
By constructing a geographic spatiotemporal relationship map and performing hierarchical iterative calculations, combined with image preprocessing and pedestrian re-identification modules, the problems of low computational efficiency and poor robustness in large-scale surveillance networks are solved, achieving efficient and accurate tracking of pedestrian paths.
Patent Information
- Application Number
- CN202511041370.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-28
- Publication Date
- 2025-11-18
AI Technical Summary
Existing technologies are computationally inefficient when processing massive amounts of video data in large-scale surveillance networks. Pre-trained models are not robust enough in complex environments and have insufficient applicability, leading to misjudgments and omissions in pedestrian path tracking. Furthermore, traditional methods are difficult to describe the movement path logic of pedestrians between multiple cameras.
By constructing a geographic spatiotemporal relationship map, performing hierarchical iterative calculations, combining an image preprocessing module for image restoration and enhancement, using a pedestrian re-identification module for comparison, clarifying pedestrian path logic, and employing the YOLOv8 framework and DeepSort tracking method for target detection and tracking.
It improves computational efficiency, enhances the robustness of the model in complex environments, meets the needs of diverse scenarios, clearly describes the pedestrian path logic, reduces the false recognition rate, and improves the accuracy of pedestrian re-identification.
Smart Images
Figure CN120976876A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision, and more particularly to a pedestrian re-identification method based on path tracking. Background Technology
[0002] With the acceleration of urbanization and continuous population growth, public safety, traffic management, and other fields face severe challenges. For example, after a public safety incident, searching through surveillance footage relies solely on manual retrieval, resulting in a massive workload and significant time waste. Against the backdrop of rapid development in computer vision technology, pedestrian detection and target tracking based on the YOLOv8 framework have been widely applied in traffic safety, park security, and autonomous driving. Image processing technologies such as super-resolution enhancement and viewpoint synthesis are beginning to be used in image restoration and AI inference. However, how to efficiently deploy and accurately identify and track pedestrians in complex and unfamiliar environments still faces many challenges.
[0003] Massive amount of detection data: In large-scale surveillance networks, there are numerous cameras, generating a massive amount of video data. Traditional algorithms struggle to efficiently process such large-scale data, resulting in high computational resource consumption and low detection efficiency.
[0004] Robustness of pre-trained models: Different application scenarios, such as lighting, weather, and shooting angle, will affect pedestrian images. Existing pre-trained models are difficult to maintain stable recognition performance in complex and ever-changing real-world environments. Their recognition accuracy cannot reach the effect of training and testing on open-source datasets, and they are prone to misjudgment and missed judgment.
[0005] High retraining costs: Different application scenarios involve a huge workload for data collection, labeling, and model training. At the same time, it is necessary to adjust the pedestrian re-identification algorithm strategy and improve the rank 1 hit rate regardless of cost. However, the existing advanced algorithms are far from meeting the requirement of 100% hit rate, so the time cost is difficult to estimate.
[0006] Poor applicability: Different application scenarios have different requirements for pedestrian re-identification. For example, shopping malls focus on pedestrians' shopping behavior paths, while transportation hubs focus on pedestrians' flow direction. General pedestrian re-identification methods are difficult to meet the diverse needs of different scenarios.
[0007] The logic of pedestrian path tracking is difficult to describe: Traditional methods are not intuitive and clear enough in describing the logic of pedestrian movement between multiple cameras, which is not conducive to understanding and application. Summary of the Invention
[0008] The main technical problem this invention aims to solve is how to address the issue of massive amounts of detection data making it difficult to meet computational efficiency requirements, and the low robustness, poor applicability, and susceptibility to misjudgments leading to pedestrian path tracking errors in unfamiliar application scenarios using existing advanced algorithms' pre-trained models. To address these problems, this invention proposes a pedestrian re-identification method based on path tracking.
[0009] The present invention adopts the following technical solution:
[0010] This invention includes the following steps:
[0011] S1. Construct a geographic spatiotemporal relationship map based on the deployment location of the camera group, determine the camera group that first captures the target pedestrian as the current round camera group, and add it to the path record set;
[0012] S2. Determine the current round of camera groups as the query camera groups, and select all neighboring camera groups of the current round of camera groups as the image library camera groups based on the geographic spatiotemporal relationship map.
[0013] S3. Input the surveillance videos captured by the query camera group and the image library camera group into the image preprocessing module to extract pedestrian images as the query set and the image library set respectively. Then, input the query set and the image library set into the pedestrian re-identification module for comparison to obtain several hit images.
[0014] S4. Judge all hit images:
[0015] If a target pedestrian is found in the hit image, the camera that captured the current hit image is added to the path record set, and then the camera that captured the current hit image is set as the camera group for the current round and the process returns to step S2.
[0016] If the target pedestrian is not found in the image, the path trajectory of the target pedestrian is output based on the path record set.
[0017] The pedestrian geographic spatiotemporal relationship map includes road information, building entrances and exits, preset key locations, and camera deployment locations.
[0018] The camera groups in the path recording set are arranged in the order they were added to the path recording set, thereby constructing the path trajectory of the target pedestrian.
[0019] The specific steps in step S3 for querying the surveillance videos captured by the camera group and the gallery camera group are as follows: select the surveillance videos captured during the required time period by combining the spatiotemporal geographic relationship map.
[0020] In step S3, the image preprocessing module extracts pedestrian images by performing target detection on the surveillance videos captured by the query camera group and the image library camera group to obtain all pedestrian images. Then, image restoration and enhancement, normalization and data enhancement are performed on all pedestrian images in sequence to construct the query set and the image library set.
[0021] The hit image is specifically a pedestrian image that is compared and determined to be the same pedestrian by the pedestrian re-identification module from pedestrian images in the image library and the query set.
[0022] In step S4, if at least one of the hit images contains a target pedestrian, then it is determined that the hit image contains a target pedestrian; otherwise, it is determined that the hit image does not contain a target pedestrian.
[0023] The beneficial effects of this invention are as follows:
[0024] This invention employs a hierarchical and iterative calculation strategy for camera groups, enabling the orderly processing of video data generated by a large number of cameras, thereby improving computational efficiency and reducing computational resource consumption.
[0025] This invention constructs a pedestrian geographic spatiotemporal relationship map based on different application scenarios, which can flexibly adapt to diverse scenario requirements and meet the requirements for pedestrian path tracking in different scenarios. Through clear steps and hierarchical iterative calculations, it can clearly and intuitively describe the movement path logic of pedestrians between multiple cameras, making it easy to understand and apply.
[0026] In the solution of this invention, the image preprocessing module employs various techniques to perform image restoration, enhancement, normalization, and data augmentation, such as super-resolution enhancement and viewpoint synthesis, enabling the pedestrian re-identification module to maintain a high recognition hit rate in complex real-world environments (different lighting, weather, shooting angles, etc.), thereby enhancing the robustness of the pre-trained model. Attached Figure Description
[0027] Figure 1 This is a schematic diagram of the steps of the present invention;
[0028] Figure 2 This is a pedestrian geographic spatiotemporal information relationship diagram according to an embodiment of the present invention;
[0029] Figure 3 This is a schematic diagram illustrating the effect of the pedestrian target detection + DeepSort tracking method based on YOLOv8 in an embodiment of the present invention;
[0030] Figure 4 This is a schematic diagram illustrating the iterative process and target pedestrian movement path of an embodiment of the present invention. Detailed Implementation
[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail, clearly, and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0032] like Figure 1 As shown, the pedestrian re-identification method in this embodiment includes the following steps:
[0033] S1. Construct a geographic spatiotemporal relationship map based on the deployment location of the camera group and the spatial characteristics of the intersection monitored by the camera group. This map is used to determine the adjacency relationship between each camera group, identify the camera group where the target pedestrian first appears as the current round camera group, and add it to the path record set.
[0034] In the pedestrian path tracking application scenario of the campus teaching area in this embodiment, the core objective is to achieve rapid and accurate localization and re-identification of target pedestrians by analyzing video sets, thereby obtaining the geographic and spatiotemporal path of the target pedestrians. Figure 2 As shown, in this embodiment, based on the satellite image, traffic map, and monitoring deployment map of the application scenario, the location of each monitored intersection that a pedestrian may pass through is marked, and a pedestrian geographic spatiotemporal relationship map is constructed to realize the binding of multiple sets of information such as the map, intersections, monitoring videos, and spatial locations. Figure 2 In the pedestrian geographic spatiotemporal relationship diagram, the circular symbol En represents a surveillance camera node, and the solid line fn with a double-headed arrow represents a pedestrian movement path. There are a total of 18 nodes and 30 movement paths in this application scenario.
[0035] S2. Determine the current round of camera groups as the query camera groups, and select all camera groups adjacent to the current round of cameras as the image library camera groups based on the geographic spatiotemporal relationship map.
[0036] S3. Input the surveillance videos captured by the query camera group and the image library camera group into the image preprocessing module to extract pedestrian images as the query set and the image library set respectively. Then, input the query set and the image library set into the pedestrian re-identification module for comparison to obtain several hit images and record the camera group corresponding to each hit image.
[0037] This involves collecting surveillance videos from nearby camera groups An and Bn in a selected geographic and spatiotemporal context, and constructing application scenario video sets hierarchically based on confirmation and elimination mechanisms and iterative calculation strategies. The application scenario video sets are collections of videos taken to construct query sets and image library sets.
[0038] Taking this embodiment as an example, based on the pedestrian geospatial-geographic relationship map, surveillance cameras are set up at 18 nodes. The surveillance videos of the 18 nodes are selected according to the time period for pedestrian path tracking analysis, and the recorded data within that time period is downloaded. First, it is confirmed that the target pedestrian appears in the image of monitoring node E1. E1 is set as level A, serving as the query set data source. Nodes E2 and E18, adjacent to node E1, are set as level B, serving as the image library data source. Subsequently, if the target pedestrian is hit by the re-identification model in node E2, node E18 is removed, and the target pedestrian image in E2 is confirmed as the query set. Nodes E18, E17, E16, and E3, adjacent to node E2, are set as level C, serving as the image library data source, and so on, iteratively calculating.
[0039] The application scenario video set is then imported into the image preprocessing module to obtain a query set and image library set that conform to the pre-trained model of the pedestrian re-identification module. The pedestrian re-identification module then detects the dataset to obtain the hit images and pedestrian identity information.
[0040] like Figure 3 As shown, in this embodiment, the video from the data source is first input into the YOLOv8 framework and the DeepSort motion trajectory tracking algorithm for target detection, unique identification, pedestrian labeling, and exporting pedestrian segmentation images. Super-resolution enhancement technology is used to repair and normalize the segmented images, and viewpoint synthesis technology is used to enhance the different viewpoint information of pedestrians in the images. The purpose is to enhance the image data in the query set and image library in unfamiliar application scenarios, improve the robustness of recognition, and make it easier for the model to match images from other viewpoints in the two sets. Finally, a query set and image library set conforming to the data format of the pedestrian re-identification module are obtained. Multiple technologies are combined to process the original video data, improving image quality and data usability. Finally, the dataset is imported into the pedestrian re-identification module for detection, obtaining hit images with high similarity to the target pedestrian images in the query set and pedestrian identity information.
[0041] YOLOv8, Deepsort, super-resolution enhancement, and viewpoint synthesis techniques can be any advanced method, requiring only modifications to their data interfaces. The person re-identification module can also employ any advanced method, including fine-tuning, pre-trained models, or retrained models. While mAP (average accuracy per second) is not considered, a high rank-10 hit rate is guaranteed, and the system can export the hit images and pedestrian identification information. This approach provides flexibility in technique selection while ensuring accuracy in person re-identification.
[0042] S4. Judge all hit images:
[0043] If a target pedestrian is found in the hit image, the hit image is retained and the camera that captured the current hit image is added to the path record set. Then, the camera that captured the current hit image is set as the camera group for the current round and the process returns to step S2.
[0044] If the target pedestrian is not found in the image, the path trajectory of the target pedestrian is output based on the path record set.
[0045] Figure 1 The text connected by the dotted lines is a functional description of the method steps, not a description of two branches being processed in parallel.
[0046] Specifically, the process involves determining whether a target pedestrian exists in the hit image: if a target pedestrian exists, invalid hit images and their corresponding cameras Bx are removed based on the pedestrian's identity information in the hit image, the geographic and spatiotemporal information of the valid hit images and camera B is confirmed, and Cn related to the path of B is selected as the data source for the image set, and the next iteration calculation is performed; if no target pedestrian exists, the iteration ends and the pedestrian movement path AB-… is output.
[0047] After each path is iterated, since even the most advanced algorithms cannot avoid issues such as missed detections and false detections, a check is performed to compare the calculated hit image with the target pedestrian in the query set. If the target pedestrian is found, the camera group is iterated at each level before proceeding to the next iteration. If the target pedestrian is not found, the target pedestrian's movement path AB-… is directly output. This directly prevents the re-identification module from misjudging the hit image due to factors such as visual environment, changing clothes, or occlusion, thus improving the accuracy of pedestrian path tracking.
[0048] like Figure 4 As shown, in this embodiment, when the target pedestrian finishes the first iteration calculation, the E2+E18 dataset obtains the hit image rank10. All 9 images in rank10 are provided by the E2 node. The images hit by the E18 node are invalid images. Therefore, the E18 node is removed, and the E18, E17, E16 and E3 nodes adjacent to the E2 node are confirmed to be set to level C. The nodes are recorded as E1-E2, and then the next iteration calculation is entered.
[0049] At the end of the second iteration, the E18+E17+E16+E3 dataset yielded the rank 10 of the hit images. All 10 images in rank 10 were provided by the E3 node, and the other nodes did not have any hit images. Therefore, the remaining nodes were removed, and the E15, E14, E4, E5, and E6 nodes adjacent to the E3 node were set to level D. The nodes were recorded as E1-E2-E3, and then the next iteration was started.
[0050] At the end of the third iteration, the E15+E14+E4+E5+E6 dataset yielded the hit image rank10. Three images in rank10 were provided by node E4, four by node E5, and three by node E6. Therefore, nodes E15 and E14 were removed. Based on the time information of the hit images, the latest time record was E6. Thus, the pedestrian path was: E1-E2-E3-E4-E5-E6. It was confirmed that E7, which is adjacent to node E6, is of level E, and then the next iteration calculation was performed.
[0051] This process continues until the target pedestrian no longer appears in the hit image when node E18 is used as the query set. Record the target pedestrian's movement nodes: E1-E2-E3-E4-E5-E6-E7-E8-E9-E10-E11-E14-E17-E18, and output the target pedestrian's movement path: f2-f6-f10-f13-f14-f16-f18-f19-f20-f23-f24-f27-f30.
[0052] Through the above embodiments, the present invention can effectively track pedestrian movement paths in practical application scenarios, greatly reduce the amount of computation and the false recognition rate, and the advanced algorithms in the image preprocessing module and pedestrian re-identification module can be replaced, thus solving the applicability of pedestrian re-identification.
[0053] The pedestrian geographic spatiotemporal relationship map includes traffic maps, all road information, building entrances and exits, preset key locations and camera deployment locations in the application scenario, and pre-marks the location of each monitored intersection that pedestrians may pass through, realizing the binding of multiple sets of information such as map, intersection, monitoring video, and spatial location.
[0054] The key locations pre-defined include entrances and exits of enclosed spaces, roads with high traffic volume, intersections or corners of major thoroughfares, and entrances to specific functional areas (such as shops, office areas, and parking lot entrances and exits).
[0055] When the application scenario is a closed space, the camera group covering all entrances and exits is selected as the candidate set of camera groups where the target pedestrian first appears, based on the geographic spatiotemporal relationship map, and the camera group where the target pedestrian first appears is determined from it.
[0056] The camera groups in the path recording set are arranged in the order they were added to the path recording set, thereby constructing the path trajectory of the target pedestrian.
[0057] In step S3, the surveillance videos captured by the camera group and the gallery camera group are queried, and the surveillance videos captured in the required time period are selected based on the spatiotemporal geographic relationship map.
[0058] Specifically, based on the spatiotemporal geographic relationship map, the location, parameters, video data format, and time period to be calculated are set to clarify the specific information to be collected and ensure that the collected video data meets the requirements for subsequent processing.
[0059] In step S3, the image preprocessing module extracts the target pedestrian images as follows: using the YOLOv8 framework and the Deepsort motion trajectory tracking algorithm, target detection is performed on the surveillance videos captured by the query camera group and the image library camera group to obtain all pedestrian images. Then, image inpainting enhancement, normalization, and data augmentation are performed on all pedestrian images in sequence to obtain processed pedestrian images. Finally, the query set and image library set are constructed based on the processed pedestrian images.
[0060] The image preprocessing module uses super-resolution enhancement technology to inpaint, enhance, and normalize images, and then uses viewpoint synthesis technology to enhance data.
[0061] The hit images are specifically pedestrian images from the image library and the query set, which are input into the pedestrian re-identification module for comparison and are determined to be images of the same pedestrian.
[0062] In step S4, if at least one of the hit images contains the target pedestrian, then it is determined that the hit image contains the target pedestrian; otherwise, it is determined that the hit image does not contain the target pedestrian.
[0063] The above is a detailed description of specific embodiments of the present invention. However, the present invention is not limited to the specific embodiments described above. Those skilled in the art can make various modifications or alterations to the above technical solutions without affecting the substantive content of the present invention, and all such modifications or alterations should be within the protection scope of the present invention.
[0064] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A pedestrian re-identification method based on path tracking, characterized in that, Includes the following steps: S1. Construct a geographic spatiotemporal relationship map based on the deployment location of the camera group, determine the camera group that first captures the target pedestrian as the current round camera group, and add it to the path record set; S2. Determine the current round of camera groups as the query camera groups, and select all neighboring camera groups of the current round of camera groups as the image library camera groups based on the geographic spatiotemporal relationship map. S3. Input the surveillance videos captured by the query camera group and the image library camera group into the image preprocessing module to extract pedestrian images as the query set and the image library set respectively. Then, input the query set and the image library set into the pedestrian re-identification module for comparison to obtain several hit images. S4. Judge all hit images: If a target pedestrian is found in the hit image, the camera that captured the current hit image is added to the path record set, and then the camera that captured the current hit image is set as the camera group for the current round and the process returns to step S2. If the target pedestrian is not found in the image, the path trajectory of the target pedestrian is output based on the path record set.
2. The pedestrian re-identification method based on path tracking according to claim 1, characterized in that: The pedestrian geographic spatiotemporal relationship map includes road information, building entrances and exits, preset key locations, and camera deployment locations.
3. The pedestrian re-identification method based on path tracking according to claim 1, characterized in that: The camera groups in the path recording set are arranged in the order they were added to the path recording set, thereby constructing the path trajectory of the target pedestrian.
4. The pedestrian re-identification method based on path tracking according to claim 1, characterized in that: The specific steps in step S3 for querying the surveillance videos captured by the camera group and the gallery camera group are as follows: select the surveillance videos captured during the required time period by combining the spatiotemporal geographic relationship map.
5. The pedestrian re-identification method based on path tracking according to claim 1, characterized in that: In step S3, the image preprocessing module extracts pedestrian images by performing target detection on the surveillance videos captured by the query camera group and the image library camera group to obtain all pedestrian images. Then, image restoration and enhancement, normalization and data enhancement are performed on all pedestrian images in sequence to construct the query set and the image library set.
6. The pedestrian re-identification method based on path tracking according to claim 1, characterized in that: The hit image is specifically a pedestrian image that is compared and determined to be the same pedestrian by the pedestrian re-identification module from pedestrian images in the image library and the query set.
7. The pedestrian re-identification method based on path tracking according to claim 1, characterized in that: In step S4, if at least one of the hit images contains a target pedestrian, then it is determined that a target pedestrian exists in the hit image. Otherwise, it is determined that there is no target pedestrian in the hit image.