Indoor cross-camera target tracking method combined with video space coordinate system conversion

Abstract

The application discloses an indoor cross-camera target tracking method combined with video space coordinate system conversion, belongs to the technical field of computer vision and intelligent monitoring, and aims at the problems of single-camera visual field limitation, insufficient multi-camera cooperation, insufficient pedestrian re-identification technology precision restricting application effect, limited hardware positioning system applicability and high cost, the method comprises the following steps: constructing a video and space coordinate system visual mapping management module, solving a conversion matrix through a perspective transformation algorithm, realizing accurate mapping of video coordinates to space coordinates, establishing a unified space coordinate system of multiple cameras, deploying a neural network human body detection and target tracking module, adopting a YOLO target tracking technology stack, and extracting structured trajectory data of targets in each video; the application realizes multi-camera target trajectory positioning by innovatively accurately mapping a video coordinate system and a space coordinate system and combining target detection and re-identification technology.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and intelligent monitoring technology, specifically involving an indoor cross-camera target tracking method that combines video spatial coordinate system transformation. Background Technology

[0002] With the rapid development of intelligent technologies, target trajectory positioning technology based on video analytics has shown broad application prospects in many fields such as security monitoring, commercial customer flow analysis, and smart building management. In commercial scenarios, precise analysis of customer movement trajectories can provide in-depth insights into consumer behavior characteristics, offering data support for store layout optimization, merchandise display improvement, and marketing strategy formulation. In the security field, continuous personnel tracking can effectively improve the level of regional security management, providing technical support for abnormal behavior early warning and key personnel monitoring. In smart building management, accurate trajectory positioning helps optimize space utilization and improve venue operational efficiency.

[0003] However, despite the urgent application demand and broad market prospects, existing technologies still face significant technical bottlenecks in achieving high-precision, continuous trajectory positioning. Especially in practical applications, issues such as trajectory breaks, misidentification, and insufficient positioning accuracy caused by technological limitations severely restrict the large-scale commercial application of these technologies. These problems not only affect the accuracy and reliability of data analysis but also reduce the practical value of the system in real-world scenarios. Overcoming these technological limitations to achieve stable and reliable high-precision continuous trajectory positioning has become a key technical challenge that urgently needs to be addressed in this field. The following three common problems also exist in pedestrian trajectory analysis: 1. Limitations of single-camera field of view and insufficient multi-camera collaboration: The fixed angle and limited coverage of a single camera prevent it from fully capturing the motion trajectory of a target within a complex scene. When a target moves between the fields of view of different cameras, the lack of an effective cross-camera correlation mechanism makes it difficult for the system to continuously track the trajectory. Although existing solutions expand the monitoring range by increasing the number of cameras, each camera often operates independently, forming information silos and failing to provide complete trajectory data support for applications such as commercial customer flow analysis. 2. Insufficient accuracy of pedestrian re-identification technology limits its application effectiveness: Traditional visual feature-based re-identification technology performs poorly in complex indoor environments. Commercial scenarios are characterized by dense crowds, complex lighting conditions, and potentially similar customer attire, leading to a high false alarm rate. Existing algorithms rely excessively on appearance features and lack comprehensive utilization of target spatial location information, making it difficult to meet the high accuracy requirements of commercial analysis and limiting its application value in refined management such as passenger flow statistics and heat map analysis. 3. Hardware positioning systems have limited applicability and are costly: While hardware positioning solutions based on signals such as Wi-Fi and Bluetooth can provide continuous location information, they require the deployment of dedicated equipment and infrastructure, resulting in high implementation costs and significant modification difficulties. In commercial scenarios, requiring customers to wear dedicated devices is clearly impractical, while methods relying solely on the MAC address of customers' mobile phone signals suffer from low coverage and unstable accuracy, hindering large-scale application.

[0004] Therefore, it is necessary to combine video spatial coordinate system transformation with indoor cross-camera target tracking methods to solve the problems of limited field of view of single cameras and insufficient multi-camera collaboration in existing technologies, insufficient accuracy of pedestrian re-identification technology which restricts application effect, and limited applicability and high cost of hardware positioning systems. Summary of the Invention

[0005] The purpose of this invention is to provide an indoor cross-camera target tracking method that incorporates video spatial coordinate system transformation, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an indoor cross-camera target tracking method combining video spatial coordinate system transformation, comprising: S1. Construct a video and spatial coordinate system visualization mapping management module, solve the transformation matrix through perspective transformation algorithm, realize accurate mapping from video coordinates to spatial coordinates, and establish a unified spatial coordinate system for multiple cameras; S2. Deploy a neural network human detection and target tracking module, using the YOLOv11+DeepSORT technology stack to extract structured trajectory data of targets in each video stream, including target ID, timestamp, video coordinates and appearance feature vectors; S3. By using the fusion algorithm of target spatial coordinates and target Re-ID, a dual-verification fusion deduplication framework is constructed, and the weights of spatial coordinates and Re-ID features are dynamically allocated to achieve the merging and continuous tracking of target trajectories across cameras.

[0007] It should be noted in the solution that the video and spatial coordinate system visualization mapping management module described in S1 includes: S101, Perspective Transformation Matrix Solving: The system provides a visual interactive interface, allowing users to establish a mapping relationship between video footage and spatial planar diagrams by dragging and dropping. S102: Multi-channel video space is unified and integrated. Based on the actual scene CAD plan, a 1:1 scale spatial coordinate system is established, and all cameras are mapped to this unified coordinate system. The overlapping areas of the field of view are processed through a weighted fusion algorithm to ensure the coordinate continuity of the target when switching between cameras. It supports dynamic parameter adjustment and has real-time mapping quality monitoring and anomaly detection capabilities. S103. Data storage and interface design, including the use of standardized JSON format to store mapping parameters and accuracy indicators, including camera configuration, spatial coordinate system, calibration point information and accuracy evaluation data; providing real-time coordinate transformation API, batch processing interface and quality monitoring interface, and realizing data association with human detection module through timestamp and camera ID.

[0008] It is further worth noting that the perspective transformation matrix solution described in S101 includes the following expression for the transformation matrix: The transformation matrix parameters are solved using least squares optimization, and the formula is as follows: .

[0009] It should be further noted that each video channel in S101 is configured with at least 4 non-collinear mapping points, and more than 16 mapping points can achieve decimeter-level mapping accuracy.

[0010] As a preferred implementation, the neural network human detection and target tracking module deployed in S2 includes: S201. Using the YOLOv11+DeepSORT technology stack, an end-to-end multi-target detection and tracking system is built. S202, Each video stream runs an independent YOLOv11+DeepSORT detection and tracing pipeline; S203. Real-time output of structured trajectory data containing target ID, timestamp, video coordinates and appearance feature vectors, which is transmitted to subsequent modules through a standardized interface.

[0011] As a preferred implementation, the DeepSORT algorithm described in S201 predicts the target motion state based on Kalman filtering, and the state vector contains information such as position, velocity, and scale; it also uses the OSNet network to extract a 512-dimensional feature vector, and uses appearance feature similarity calculation to assist in data association; and it integrates IoU matching, appearance feature matching, and motion consistency matching to construct a multi-dimensional association cost matrix.

[0012] As a preferred implementation, the dual-verification fusion deduplication framework constructed in S3 through the fusion algorithm of target spatial coordinates and target Re-ID, and the dynamic allocation of weights for spatial coordinates and Re-ID features, includes: S301. Based on the complementary characteristics of spatial coordinates and Re-ID visual features, a dual-verification fusion deduplication framework is constructed; spatial coordinates provide target physical location information, and Re-ID features provide target visual identity information. The two mutually verify each other to improve the accuracy of cross-camera target association. S302. Establish a collaborative verification mechanism for spatial coordinates and Re-ID features, and achieve optimal matching results through dynamic weight allocation. The weight model expression is as follows: ,in, For scene target density, For light quality, To adjust the parameters; S303, trajectory complementarity and deduplication, spatial coordinates provide spatial constraints for Re-ID matching, narrowing the search range; Re-ID features provide identity verification for targets in similar spatial locations, preventing false associations.

[0013] As a preferred implementation, the weight allocation rule in S302 is as follows: In sparse scenarios, spatial coordinates have a higher weight, and target association is mainly based on location information. In dense scenes, the Re-ID feature weights are increased, and similar targets are distinguished based on appearance features; When lighting conditions are good, increase feature weights to make full use of visual information; When lighting conditions are poor, spatial weights are increased to ensure the reliability of basic matching; The weight allocation adopts an automatic adjustment method based on scene quality assessment to ensure that the best performance balance is maintained in various environments.

[0014] As a preferred implementation, in S303, when different cameras detect targets that are spatially close, they first compare the spatial coordinate distances; if the distance is less than a set threshold, they calculate the Re-ID feature similarity. A comprehensive matching score is then generated by weighted fusion of spatial distance and feature similarity, using the following formula: Match score = λ × spatial similarity + (1-λ) × feature similarity Spatial similarity is calculated based on Euclidean distance, while feature similarity is measured by cosine distance. When the overall matching score exceeds a predetermined threshold, the system determines that they are the same target and performs trajectory merging and identity deduplication.

[0015] Compared with existing technologies, the indoor cross-camera target tracking method combining video spatial coordinate system transformation provided by this invention has at least the following beneficial effects: By innovatively mapping video coordinates to spatial coordinates and combining target detection and re-identification technologies, target trajectory localization across multiple cameras is achieved. First, by utilizing multi-camera collaboration and spatial coordinate constraints, the limitations of single-camera field of view are effectively overcome, enabling continuous and accurate target localization in physical space. Second, by fusing spatial information and visual features, the reliability of target re-identification is significantly improved, reducing misjudgments caused by appearance changes or occlusion. Finally, the solution fully utilizes existing surveillance infrastructure, eliminating the need for dedicated hardware deployment, and features flexible deployment and low cost, making it suitable for various scenarios ranging from small retail stores to large commercial spaces. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the video and spatial coordinate system visualization interactive interface of the present invention; Figure 2 This is a schematic diagram of the indoor target trajectory of the present invention; Figure 3 This is a schematic diagram of the multi-camera tracking system of the present invention. Detailed Implementation

[0017] The present invention will be further described below with reference to embodiments.

[0018] This invention provides an indoor cross-camera target tracking method that incorporates video spatial coordinate system transformation, comprising the following steps: S1. Construct a video and spatial coordinate system visualization mapping management module, solve the transformation matrix through perspective transformation algorithm, realize accurate mapping from video coordinates to spatial coordinates, and establish a unified spatial coordinate system for multiple cameras; S101. Perspective Transformation Matrix Solving: The system provides a visual interactive interface, allowing users to establish mapping relationships between video frames and spatial plans through dragging and dropping. Each video stream should have at least 4 non-collinear mapping points, with 8 points recommended. More than 16 mapping points can achieve decimeter-level accuracy. A perspective transformation algorithm is used, and the transformation matrix is ​​optimized using the least squares method. The expression for the transformation matrix is: The transformation matrix parameters are solved using least squares optimization, and the formula is as follows: ; Reference Figure 1 As shown, the system calculates the mapping error in real time and provides an error heatmap to guide users in optimizing the distribution of points. S102: Multi-channel video space is unified and integrated. Based on the actual scene CAD plan, a 1:1 scale spatial coordinate system is established, and all cameras are mapped to this unified coordinate system. The overlapping areas of the field of view are processed through a weighted fusion algorithm to ensure the coordinate continuity of the target when switching between cameras. It supports dynamic parameter adjustment and has real-time mapping quality monitoring and anomaly detection capabilities. S103. Data storage and interface design, including the use of standardized JSON format to store mapping parameters and accuracy indicators, including camera configuration, spatial coordinate system, calibration point information and accuracy evaluation data; providing real-time coordinate transformation API, batch processing interface and quality monitoring interface, and realizing data association with human detection module through timestamp and camera ID; S2. Deploy a neural network human detection and target tracking module, using the YOLOv11+DeepSORT technology stack to extract structured trajectory data of targets in each video stream, including target ID, timestamp, video coordinates and appearance feature vectors; S201. An end-to-end multi-target detection and tracking system is constructed using the YOLOv11+DeepSORT technology line. YOLOv11 is responsible for real-time target detection, while DeepSORT achieves robust multi-target tracking. The two work together to extract target trajectory data from each video stream. DeepSORT predicts the target motion state based on Kalman filtering, and the state vector contains information such as position, velocity, and scale. The re-identification module uses the OSNet network to extract a 512-dimensional feature vector and assists in data association through appearance feature similarity calculation. A multi-dimensional association cost matrix is ​​constructed by integrating IoU matching, appearance feature matching, and motion consistency matching. S202. Each video stream runs the YOLOv11+DeepSORT detection and tracking pipeline independently. The YOLO model is responsible for detecting human targets from video frames and outputting target bounding boxes and confidence scores. The DeepSORT module associates targets based on the detection results, assigns a unique ID to each target and generates continuous trajectories. S203. Real-time output of structured trajectory data containing target ID, timestamp, video coordinates and appearance feature vectors, which is transmitted to subsequent modules through a standardized interface; As the front-end processing unit of the system, this module forms a complete technical chain with modules such as spatial coordinate mapping and cross-camera trajectory association. By optimizing algorithm parameters, it ensures efficient and stable operation in multi-video scenarios, and ultimately achieves accurate multi-camera target trajectory positioning function. S3. A dual-verification fusion and deduplication framework is constructed using a fusion algorithm that integrates target spatial coordinates and target Re-ID features. This framework dynamically allocates weights for spatial coordinates and Re-ID features, referencing... Figure 2 As shown, this enables the merging and continuous tracking of target trajectories across cameras; S301. Based on the complementary characteristics of spatial coordinates and Re-ID visual features, a dual-verification fusion deduplication framework is constructed; spatial coordinates provide target physical location information, and Re-ID features provide target visual identity information. The two mutually verify each other to improve the accuracy of cross-camera target association. S302. Establish a collaborative verification mechanism for spatial coordinates and Re-ID features, and achieve optimal matching results through dynamic weight allocation. The weight model expression is as follows: , For scene target density, For light quality, To adjust the parameters, The weighting rules are as follows: In sparse scenarios, spatial coordinates have a higher weight, and target association is mainly based on location information. In dense scenes, the Re-ID feature weights are increased, and similar targets are distinguished based on appearance features; When lighting conditions are good, increase feature weights to make full use of visual information; When lighting conditions are poor, spatial weights are increased to ensure the reliability of basic matching; The weight allocation adopts an automatic adjustment method based on scene quality assessment to ensure that the best performance balance is maintained in various environments; S303, trajectory complementarity and deduplication, spatial coordinates provide spatial constraints for Re-ID matching, narrowing the search range; Re-ID features provide identity verification for targets with similar spatial locations, preventing false associations; Reference Figure 3 As shown, when different cameras detect targets that are spatially close, they first compare the spatial coordinate distances. If the distance is less than a set threshold, Re-ID feature similarity is calculated. A comprehensive matching score is generated by weighted fusion of spatial distance and feature similarity, using the following formula: Match score = λ × spatial similarity + (1-λ) × feature similarity Spatial similarity is calculated based on Euclidean distance, while feature similarity is measured by cosine distance. When the overall matching score exceeds a predetermined threshold, the system determines that they are the same target and performs trajectory merging and identity deduplication.

[0019] Example A 25-square-meter unmanned supermarket was selected as the implementation scenario, and three cameras were deployed to cover the entire scene. The following metrics were obtained:

[0020] In summary, the advantages of this invention are as follows: By innovatively and accurately mapping the video coordinate system to the spatial coordinate system, and combining target detection and re-identification technologies, target trajectory localization across multiple cameras is achieved; firstly, by utilizing multi-camera collaboration and spatial coordinate constraints, the limitation of single-camera field of view is effectively overcome, achieving continuous and accurate target localization in physical space; secondly, by fusing spatial information and visual features, the reliability of target re-identification is significantly improved, reducing misjudgments caused by appearance changes or occlusion; finally, the solution makes full use of existing monitoring infrastructure, requiring no dedicated hardware deployment, and features flexible deployment and low cost, making it suitable for various scenarios from small retail stores to large commercial spaces.

[0021] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

[0022] The scope of protection claimed by this invention is defined by the appended claims and their equivalents.

Claims

1. An indoor cross-camera target tracking method combining video spatial coordinate system transformation, characterized in that, include: S1. Construct a video and spatial coordinate system visualization mapping management module, solve the transformation matrix through perspective transformation algorithm, realize accurate mapping from video coordinates to spatial coordinates, and establish a unified spatial coordinate system for multiple cameras; S2. Deploy a neural network human detection and target tracking module, using the YOLOv11+DeepSORT technology stack to extract structured trajectory data of targets in each video stream, including target ID, timestamp, video coordinates and appearance feature vectors; S3. By using the fusion algorithm of target spatial coordinates and target Re-ID, a dual-verification fusion deduplication framework is constructed, and the weights of spatial coordinates and Re-ID features are dynamically allocated to achieve the merging and continuous tracking of target trajectories across cameras.

2. The indoor cross-camera target tracking method combining video spatial coordinate system transformation according to claim 1, characterized in that, The video and spatial coordinate system visualization mapping management module described in S1 includes: S101, Perspective Transformation Matrix Solving: The system provides a visual interactive interface, allowing users to establish a mapping relationship between video footage and spatial planar diagrams by dragging and dropping. S102: Multi-channel video space is unified and integrated. Based on the actual scene CAD plan, a 1:1 scale spatial coordinate system is established, and all cameras are mapped to this unified coordinate system. The overlapping areas of the field of view are processed through a weighted fusion algorithm to ensure the coordinate continuity of the target when switching between cameras. It supports dynamic parameter adjustment and has real-time mapping quality monitoring and anomaly detection capabilities. S103. Data storage and interface design, including the use of standardized JSON format to store mapping parameters and accuracy indicators, including camera configuration, spatial coordinate system, calibration point information and accuracy evaluation data; providing real-time coordinate transformation API, batch processing interface and quality monitoring interface, and realizing data association with human detection module through timestamp and camera ID.

3. The indoor cross-camera target tracking method combining video spatial coordinate system transformation according to claim 2, characterized in that: The perspective transformation matrix solution described in S101, wherein the transformation matrix expression is: The transformation matrix parameters are solved using least squares optimization, and the formula is as follows: .

4. The indoor cross-camera target tracking method combining video space coordinate system transformation according to claim 3, characterized in that: In S101, each video channel is configured with at least 4 non-collinear mapping points, and more than 16 mapping points can achieve decimeter-level mapping accuracy.

5. The indoor cross-camera target tracking method combining video spatial coordinate system transformation according to claim 1, characterized in that, The neural network human detection and target tracking module described in S2 includes: S201. Using the YOLOv11+DeepSORT technology stack, an end-to-end multi-target detection and tracking system is built. S202, Each video stream runs an independent YOLOv11+DeepSORT detection and tracing pipeline; S203. Real-time output of structured trajectory data containing target ID, timestamp, video coordinates and appearance feature vectors, which is transmitted to subsequent modules through a standardized interface.

6. The indoor cross-camera target tracking method combining video spatial coordinate system transformation according to claim 5, characterized in that: The DeepSORT algorithm described in S201 predicts the target motion state based on Kalman filtering, and the state vector contains information such as position, velocity, and scale. It also uses the OSNet network to extract a 512-dimensional feature vector and calculates auxiliary data association through appearance feature similarity calculation. It integrates IoU matching, appearance feature matching, and motion consistency matching to construct a multi-dimensional association cost matrix.

7. The indoor cross-camera target tracking method combining video space coordinate system transformation according to claim 1, characterized in that: The fusion algorithm described in S3, which integrates target spatial coordinates and target Re-ID, constructs a dual-verification fusion deduplication framework. The dynamic allocation of weights for spatial coordinates and Re-ID features includes: S301. Based on the complementary characteristics of spatial coordinates and Re-ID visual features, a dual-verification fusion deduplication framework is constructed; spatial coordinates provide target physical location information, and Re-ID features provide target visual identity information. The two mutually verify each other to improve the accuracy of cross-camera target association. S302. Establish a collaborative verification mechanism for spatial coordinates and Re-ID features, and achieve optimal matching results through dynamic weight allocation. The weight model expression is as follows: ,in, For scene target density, For light quality, To adjust the parameters; S303, trajectory complementarity and deduplication, spatial coordinates provide spatial constraints for Re-ID matching, narrowing the search range; Re-ID features provide identity verification for targets in similar spatial locations, preventing false associations.

8. The indoor cross-camera target tracking method combining video spatial coordinate system transformation according to claim 7, characterized in that: The weight allocation rule described in S302 is as follows: In sparse scenarios, spatial coordinates have a higher weight, and target association is mainly based on location information. In dense scenes, the Re-ID feature weights are increased, and similar targets are distinguished based on appearance features; When lighting conditions are good, increase feature weights to make full use of visual information; When lighting conditions are poor, spatial weights are increased to ensure the reliability of basic matching; The weight allocation adopts an automatic adjustment method based on scene quality assessment to ensure that the best performance balance is maintained in various environments.

9. The indoor cross-camera target tracking method combining video spatial coordinate system transformation according to claim 7, characterized in that: In S303, when different cameras detect targets that are spatially close, they first compare the spatial coordinate distances. If the distance is less than a set threshold, Re-ID feature similarity is calculated. A comprehensive matching score is generated by weighted fusion of spatial distance and feature similarity, using the following formula: Match score = λ × spatial similarity + (1-λ) × feature similarity Spatial similarity is calculated based on Euclidean distance, while feature similarity is measured by cosine distance. When the overall matching score exceeds a predetermined threshold, the system determines that they are the same target and performs trajectory merging and identity deduplication.

Images