Camera data labeling method based on UWB and target tracking

Abstract

The invention discloses a camera data labeling method based on UWB and target tracking, and belongs to the technical field of computer vision and wireless positioning, and the method mainly comprises the steps: dynamically generating a camera priority score based on the absolute three-dimensional coordinates of a target provided by the UWB in combination with motion trend prediction; selecting the camera with the highest score for scheduling according to the priority scores of the cameras in the overlapping region scheduling scene; when visual tracking fails, UWB coordinates are directly used as spatial anchor points to keep track continuity; when the vision is available, UWB positioning and visual observation are fused into a smooth track through extended Kalman filtering, and structured annotation data including the identity, the position, the timestamp and the appearance characteristics of the target are output in real time. According to the method, a real-time scheduling and labeling framework of UWB-visual depth fusion is constructed, and a high-precision and high-robustness data basis is provided for scenes such as intelligent monitoring and robot cooperation.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and wireless positioning technology, and in particular to a camera data labeling method based on UWB and target tracking. Background Art

[0002] Since digital video surveillance entered the large-scale application stage, the industry's demand for "all-weather, continuous, and uninterrupted" tracking has been growing. The earliest systems relied on a single camera for target detection and recording. Subsequently, with the maturity of GPU inference and deep learning algorithms, multi-camera collaboration and multi-object tracking (MOT) have gradually become mainstream. Large-scale scenarios (warehouses, airports, rail transit stations, and smart factories) often require twenty or even hundreds of cameras to cover the entire area. In these scenarios, targets frequently cross the overlapping fields of view of different cameras. Without a unified scheduling strategy, the system will experience track interruption or ID confusion during switching, ultimately affecting post-event retrieval, behavioral analysis, and even security decision-making. At the same time, high-precision indoor positioning technology is also rapidly evolving. Unlike GPS, which struggles to provide stable signals in obstructed environments, ultra-wideband (UWB) uses nanosecond pulse ranging to provide absolute coordinates at the centimeter to decimeter level. Therefore, it is widely used in robot navigation, asset management, and personnel positioning.

[0003] The traditional "positioning-camera" separation architecture treats UWB as a standalone service, with the camera still relying on visual features and Kalman filter tracking. The two systems don't coordinate with each other, making it difficult to ensure trajectory continuity in the face of target occlusion, sudden changes in illumination, or high-speed motion. "Multi-camera scheduling" and "UWB high-precision positioning" originally evolved within separate technical systems: the former emphasized smooth switching between different perspectives and solving cross-perspective re-identification problems, while the latter focused on absolute coordinate accuracy and mitigating non-line-of-sight (NLOS) errors and multipath interference. This results in the following flaws in existing camera data annotation: 1. Existing visual tracking technologies are prone to failure in environments with occlusion, changing lighting, or rapid target movement. The fundamental reason is that traditional algorithms (such as DeepSORT) primarily rely on association between target appearance features (color, texture) and a simplified motion model based on the Kalman filter. When a target is partially or completely obscured by other objects in the image, these appearance features disappear or severely degrade, preventing the algorithm from correctly updating the motion trajectory. Sudden changes in lighting cause drastic changes in image brightness and contrast, making it difficult for feature point detection methods such as SIFT and ORB to extract sufficiently stable visual information. High-speed target movement can also exceed the prediction or search range of the Kalman filter, causing the tracking window to deviate from the actual target position. These factors ultimately lead to problems such as gaps in the labeled data or jumps in target IDs, requiring extensive manual repair and recovery efforts, reducing overall labeling efficiency by at least half.

[0004] 2. Traditional vision methods locate targets only within the image pixel coordinate system. Obtaining the target's absolute position in the three-dimensional world often requires the use of additional depth cameras or multi-view geometry algorithms (such as triangulation). Monocular cameras cannot directly output depth information and rely on a ground plane assumption or a priori models of the human form, resulting in errors easily exceeding 20 to 30 centimeters. Using multiple cameras requires rigorous and complex calibration, which becomes ineffective if the external environment or the camera's position changes. The lack of reliable three-dimensional coordinates not only affects the accuracy of the annotation results but also limits their usefulness in high-precision applications such as robotic navigation and AR / VR spatial interaction.

[0005] 3. When attempting to simply cascade UWB positioning and visual detection, the problem of inability to deeply couple multi-source data arises. The UWB update frequency is often out of sync with the camera frame rate. Without a sophisticated time alignment mechanism, the position interpolation of moving targets will produce significant deviations. The global coordinate system output by the UWB and the local coordinate system of the camera also require calibration and conversion. If the calibration is inadequate or there are rotation matrix errors, significant coordinate jumps will occur after fusion. In multi-target scenarios, without a sufficiently comprehensive matching strategy, UWB tags and visual detection targets cannot maintain a one-to-one correspondence, making it very easy for target IDs to be confused or for tracks to cross. The resulting data jitter and position jumps can reach up to 50 to 60 centimeters, making it difficult for fused annotation to meet the needs of refined positioning and analysis.

[0006] Additionally, some existing solutions tend to use IMU sensors to compensate for visual deficiencies. However, because IMUs calculate pose through the integration of accelerometers and gyroscopes, their noise (especially bias instability) accumulates over time, causing significant deviations in the position information output by the sensor after extended operation. Typically, IMUs can generate errors approaching meters per minute, requiring continuous reliance on external methods (such as visual SLAM loop closure detection) for correction. Once labeling takes longer than half an hour to an hour, the accumulated errors further increase, significantly reducing practicality and increasing the system's computational and maintenance costs.

[0007] In summary, existing technologies face obvious deficiencies in adaptability to visual environments, depth and absolute position acquisition, spatiotemporal alignment of multi-source data, and long-term sensor accuracy, which seriously restrict the efficiency and quality of high-precision labeling. Summary of the Invention

[0008] The purpose of the present invention is to overcome the problems existing in the prior art and provide a camera data labeling method based on UWB and target tracking.

[0009] The object of the present invention is achieved through the following technical solutions: A camera data labeling method based on UWB and target tracking, comprising: Using the absolute 3D coordinates of the target provided by UWB, the coverage area of ​​each camera's field of view is calculated in real time; combined with motion trend prediction, the camera priority score is dynamically generated; The scheduling scenarios are divided into non-overlapping area scheduling and overlapping area scheduling based on the current field of view of the target. In the non-overlapping area scheduling scenario, the current camera is directly scheduled; in the overlapping area scheduling scenario, the camera with the highest score is selected for scheduling based on the camera priority score. When visual tracking fails, the UWB coordinates are directly used as spatial anchors to maintain trajectory continuity; when vision is available, the extended Kalman filter is used to integrate UWB positioning and visual observation into a smooth trajectory, and structured annotation data containing target identity, location, timestamp and appearance features are output in real time.

[0010] In some embodiments, in an overlapping area scheduling scenario, UWB coordinates are combined with visual features for cross-camera association.

[0011] In some embodiments, the cross-camera association of the combined UWB coordinates and visual features includes: First, the three-dimensional coordinates of the UWB tag are projected onto the image plane to generate a region of interest, and only the visual detection targets within the region of interest are associated; then, the Hungarian algorithm and graph neural network are combined to calculate the matching cost based on the target position, appearance features and motion consistency.

[0012] In some embodiments, the cross-camera association of the combined UWB coordinates and visual features further includes: Introduce a conflict resolution mechanism to prioritize targets with consistent movement directions.

[0013] In some embodiments, the region of interest is calculated as follows: in, Indicates the absolute three-dimensional coordinates of the target, represents the external parameter matrix of the camera, K represents the internal parameter matrix of the camera, Represents the final two-dimensional pixel position on the image.

[0014] In some embodiments, the matching cost is a function of: Among them, λ represents the weight.

[0015] In some embodiments, the absolute three-dimensional coordinates of the target provided by UWB include: In the offline phase, a calibration board with a UWB tag is deployed. By collecting multiple sets of coordinate corresponding points, the least squares method is used to solve the transformation matrix between the UWB global coordinate system and the camera coordinate system. During online operation, the calibration parameters are corrected in real time through the extended Kalman filter.

[0016] In some embodiments, dynamically generating a camera priority score in combination with motion trend prediction includes: Priority is calculated based on the target's distance from the camera and how far the target is expected to move within the camera's field of view.

[0017] In some embodiments, the priority calculation specifically includes: The method for calculating the priority score of each camera is: in, For camera The priority score, Target and camera The actual distance between Target on camera Estimated activity distance within the visual monitoring range; is the weight index, and the constraints are ; The weight of the weight index is determined by evaluating the quantitative index, and the method for determining the quantitative index is as follows: in, To evaluate quantitative indicators, and are the maximum and minimum priority scores corresponding to the cameras currently tracking the target.

[0018] In some embodiments, a detection confidence threshold is set. When the detection confidence is lower than the threshold, it is determined that visual tracking has failed and the tracking mode is automatically switched to the UWB-dominated tracking mode.

[0019] It should be further explained that the technical features corresponding to the above embodiments can be combined or replaced with each other to form a new technical solution if there is no conflict.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. Global Space Benchmark and Dynamic Priority Scheduling This method uses the absolute three-dimensional coordinates of a target, provided by UWB, to calculate in real time the coverage of each camera's field of view (e.g., distance from the target to the edge of the field of view, imaging resolution). Combined with motion trend predictions (e.g., speed and direction), it dynamically generates camera priority scores. When a target moves from the center of camera A's field of view toward its edge, the system pre-activates adjacent camera B, implementing "predictive" scheduling and avoiding track interruptions caused by switching delays. In areas where coordinates overlap, the optimal camera is selected based on imaging clarity (e.g., target pixel size), rather than relying on static trigger lines.

[0021] 2. Cross-modal ID binding and trajectory continuity assurance When an object enters an overlapping area, UWB coordinates are combined with visual features (such as Re-ID vectors) for cross-camera correlation, avoiding ID jumps caused by appearance changes in pure visual matching. When visual tracking fails due to occlusion, UWB coordinates are directly used as spatial anchors to maintain trajectory continuity, rather than relying on interpolation or manual repair.

[0022] 3. Semantic-spatial joint annotation output The system fuses visually detected category and pose information with UWB-calibrated 3D coordinates to generate structured data containing the target's identity, location, timestamp, and appearance characteristics. For example, in a smart factory scenario, the system can simultaneously output annotation information such as "Worker A was at (3.2m, 5.1m, 0m) at 10:00:00, holding tool B," providing multi-dimensional input for behavioral analysis and automated scheduling. The annotation results, which include the target's category (semantic information), 2D bounding box (visual information), and 3D coordinates (spatial location information), can be widely used in behavioral analysis, trajectory prediction, AR / VR scenarios, and various applications requiring high-precision location and semantic data.

[0023] 4. Combine UWB's absolute positioning capabilities with the semantic perception capabilities of vision to construct a joint tracking model. Specifically, a UWB tag is used to acquire the target's global 3D coordinates in real time and fuses them with the local coordinates detected by vision using an extended Kalman filter (EKF). When the target is occluded, causing the visual confidence level to fall below a threshold, the system automatically switches to UWB-dominated tracking mode. The position is predicted based on the target's motion model (e.g., a uniform velocity or uniform acceleration model) and back-projected onto the image plane to generate a virtual bounding box. Furthermore, by introducing kinematic constraints (e.g., a maximum acceleration limit), sudden trajectory changes caused by UWB multipath interference can be effectively suppressed, ensuring the smoothness of virtual annotations.

[0024] 5. The present invention proposes a phased calibration and dynamic optimization method. First, a calibration board with a UWB tag is deployed in the offline phase. By collecting multiple sets of corresponding points, the least squares method is used to solve the transformation matrix (including the rotation matrix R and the translation vector T) between the UWB global coordinate system and the camera coordinate system. During online operation, the calibration parameters are corrected in real time through the extended Kalman filter to compensate for calibration drift caused by camera vibration or ambient temperature changes. For dynamic targets, the system back-projects the pixel coordinates of visual detection into the camera coordinate system through the camera intrinsic parameter matrix, and then converts them to the UWB global coordinate system through the calibration matrix, and finally outputs structured annotation data containing three-dimensional coordinates (x, y, z). This method avoids the complex calibration requirements of traditional multi-view vision and supports spatial analysis with centimeter-level accuracy.

[0025] 6. Spatiotemporal alignment and dynamic matching of multi-source data fusion This invention uses a rigorous spatiotemporal synchronization and coordinate system calibration process to ensure that UWB data and camera data are fused at the same time and in the same reference coordinate system. During target tracking, dynamic fusion weights or Kalman filtering are used to match visual detection results with UWB positioning information, thereby improving the robustness of multi-target tracking and annotation. To address the association challenge in multi-target scenarios, a hierarchical matching strategy is proposed: first, the 3D coordinates of the UWB tag are projected onto the image plane to generate a region of interest (ROI). Only visually detected targets within the ROI are associated. The Hungarian algorithm is then combined with a graph neural network (GNN) to calculate the matching cost based on target position, appearance features, and motion consistency (such as velocity vector direction). Furthermore, the introduction of a conflict resolution mechanism (such as prioritizing association of targets with consistent motion directions) can further reduce the ID jump rate.

[0026] 7. Automated deployment and reduced manual annotation costs By means of automated spatiotemporal synchronization, coordinate alignment, and target association, the present invention reduces the tedious manual calibration or manual intervention steps, improves the efficiency of video annotation, and avoids the subjective omission and mislabeling problems that are prone to occur when manually annotating large-scale data.

[0027] 8. Enhance scenario adaptability and scalability The system can adapt to multiple target types (such as pedestrians, vehicles, robots, etc.) and supports deployment in scenarios of varying scales. Once the initial calibration is completed, it can be flexibly expanded to meet the needs of more diverse applications.

[0028] In summary, the present invention not only solves the scheduling efficiency problem of traditional multi-camera systems, but also achieves a leap from "image plane trajectory" to "real space semantics", providing a high-precision and high-robustness data foundation for scenarios such as intelligent monitoring and robot collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of a camera data labeling method based on UWB and target tracking according to an embodiment of the present invention; Figure 2 A schematic diagram illustrating the division of multi-camera scheduling scenarios according to an embodiment of the present invention; Figure 3 A schematic diagram of a target movement direction trend according to an embodiment of the present invention; Figure 4 This is a schematic diagram of overlapping areas according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The technical solutions of the present invention are described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings herein can be arranged and designed in various different configurations. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0031] It should be noted that the defects existing in the solutions in the above-mentioned prior art are the results obtained by the inventor after practice and careful research. Therefore, the discovery process of the above-mentioned problems and the solutions proposed in the embodiments of this application below for the above-mentioned problems should be the contributions made by the inventor to this application in the process of invention and creation, and should not be understood as technical contents known to technical personnel in this field.

[0032] In response to the technical problems pointed out in the background technology, the embodiments provided by the present invention are as follows: In an exemplary embodiment, a camera data annotation method based on UWB and target tracking is provided, comprising: Using the absolute 3D coordinates of the target provided by UWB, the coverage area of ​​each camera's field of view is calculated in real time; combined with motion trend prediction, the camera priority score is dynamically generated; The scheduling scenarios are divided into non-overlapping area scheduling and overlapping area scheduling based on the current field of view of the target. In the non-overlapping area scheduling scenario, the current camera is directly scheduled; in the overlapping area scheduling scenario, the camera with the highest score is selected for scheduling based on the camera priority score. When visual tracking fails, the UWB coordinates are directly used as spatial anchors to maintain trajectory continuity; when vision is available, the extended Kalman filter is used to integrate UWB positioning and visual observation into a smooth trajectory, and structured annotation data containing target identity, location, timestamp and appearance features are output in real time.

[0033] The principle of this method is: Using the absolute three-dimensional position provided by the high-frequency UWB coordinate stream as a "global prior", combined with the multi-camera coverage boundary and visual tracking results, in the process of multiple cameras tracking the target, which camera to choose to track the target is the main issue to be considered by the multi-camera scheduling algorithm. In an actual multi-camera monitoring system, the relative position of the cameras and the camera's field of view usually determine the distribution of the overlapping areas of the monitoring field of view between multiple cameras. The present invention focuses on the scheduling of multi-camera systems in a single target tracking environment. Therefore, the key problem that the scheduling algorithm needs to solve is the processing method when the target is located in the overlapping area of ​​the field of view of multiple cameras. When visual tracking fails due to occlusion or illumination, the UWB trajectory continues to maintain the target ID and three-dimensional position; when vision is available, the extended Kalman filter (EKF) is used to merge the two types of observations into a smooth trajectory, and the fused annotation of "two-dimensional box + three-dimensional coordinates + timestamp" is output in real time.

[0034] In the actual multi-camera scheduling process, the scheduling scenarios can be divided into two categories according to the current field of view of the target: non-overlapping area scheduling and overlapping area scheduling. Figure 2 In the figure, areas A and B are overlapping scheduling areas, and the rest are non-overlapping scheduling areas.

[0035] (1) Non-overlapping area scheduling scenario The current target is within the field of view of the camera, and this area does not intersect with the field of view of other cameras. At this time, the camera to be scheduled is the current camera.

[0036] (2) Overlapping area scheduling scenario To objectively and effectively evaluate the target tracking performance of multiple cameras, this paper proposes the concept of camera priority, where the priority depends on the distance between the target and the camera and the distance at which the target is expected to move within the camera's field of view. The closer the target is to the camera and the longer the distance at which the target is expected to move within the camera's field of view, the higher the priority corresponding to that camera. This can be expressed as: (1) in, represent The priority score of the camera, Representative camera The distance to the target, Represents the target expected to be in the camera The distance of activities within the field of view. and for and The corresponding weight ratio, the constraint condition is Formula (1) is simplified to: (2) The parameters directly determine the effect of the scheduling algorithm. When it is larger, the distance between the target and the camera is more important; when When it is smaller, more attention is paid to the distance at which the target is expected to move within the camera range. Therefore, in order to reasonably give different tracking scenarios the distance and expected activity distance By assigning different weights, the present invention proposes a quantitative index of scheduling effect, which is expressed as follows: (3) in, and are the maximum and minimum priority scores corresponding to the cameras currently tracking the target. The smaller the difference between the maximum and minimum priority scores of a camera, the larger the quantitative index value, indicating that the scheduling effect of the multi-camera scheduling algorithm is better.

[0037] Selected From 0 to 1, increase by 0.1 each time. That is, 0, 0.1, 0.2, 0.3, ... 0.9, 1, a total of 11 groups of selected values. Collect multiple groups of data in the overlapping field of view monitoring area (in this example, the data collected in the overlapping field of view monitoring area is 80-120 groups), and count them according to a certain time frequency. and , Substitute into formula (2), and get a total of 11 groups of data, each with multiple values. According to formula (3), calculate the evaluation index value of each group of data, and then compare the values ​​of the 11 groups of evaluation quantitative index values, and select the group with the largest evaluation quantitative index value. That is the optimal parameter we selected.

[0038] Further, let's introduce in detail and .

[0039] Representative camera Distance to target Represents the target expected to be in the camera The distance of activities within the field of view.

[0040] First, determine the target's movement direction trend. Based on the target's previous position information, draw the target's movement trajectory curve. The tangent line of the current point on the curve is the target's movement direction trend, such as Figure 3 shown.

[0041] Figure 4 is an illustration of the overlapping area. Represents the distance that the target moves within the field of view of camera 1. Represents the distance that the target moves within the field of view of camera 2. We can calculate the target location and target movement direction trend and the monitoring mapping table of camera 1 and camera 2. and size.

[0042] The introduction of UWB absolute coordinates and motion trend prediction dynamically calculates camera priority scores (coverage distance, image quality, and motion direction matching), enabling "predictive" scheduling. For example, when a target approaches the edge of camera A's field of view, camera B is activated in advance and tracked in parallel, avoiding track interruptions caused by switching delays.

[0043] Specifically, if Figure 1 As shown in Figure 2, the specific process of data annotation is given.

[0044] 1. Input tracking video stream: contains the video information of all cameras in the deployment area, obtains the location information of all cameras based on the video information, and builds a monitoring mapping table covering the boundary of the coverage area for all cameras based on the obtained information: the current position of the i-th camera and the coordinates of the coverage area boundaries ; 2. Get the location information of the tracking target: Get the location information of the current tracking target based on UWB technology ; 3. Time-space synchronization and coordinate alignment: Functional goal: to solve the time deviation and spatial coordinate system inconsistency issues between UWB and camera data. Among them, the coordinate system alignment is obtained by calibrating the camera external parameters to obtain the following matrix: , used to convert UWB world coordinates to pixel plane when needed, or to back-project pixels to world coordinates.

[0045] Timestamp alignment: The UWB base station and camera are connected to the same local area network, and the system clocks are synchronized through the NTP protocol. The time deviation is controlled within ±10ms; a high-precision software timestamp (accurate to milliseconds) is added to each frame image and UWB data packet.

[0046] Dynamic interpolation compensation: If the UWB data frequency (100 Hz) is higher than the camera frame rate (30 Hz), linear interpolation is performed on the UWB data within the time interval between camera frames to generate a coordinate sequence that matches the image timestamp.

[0047] Spatial calibration and dynamic correction: Offline calibration: Deploy a calibration board with a UWB tag in the scene and collect more than 30 sets of coordinate corresponding points; use the least squares method to solve the rotation matrix RR and translation vector TT, with an error of <5cm.

[0048] Online calibration: During operation, calibration parameters are updated every minute through feature point matching: SIFT feature points are extracted from the image and matched with the projection position of the UWB tag; outliers are removed using the RANSAC algorithm, and R and T are recalculated.

[0049] 4. Dynamic association of multiple targets: Bind UWB tags to visual detection targets one by one to avoid ID jumps.

[0050] Technical implementation: Coarse correlation (time window matching): Project the UWB coordinates to the image plane through the calibration matrix to generate the matrix ROI area: The absolute coordinates of the target in three-dimensional space (world coordinate system) measured by the UWB system.

[0051] The camera's extrinsic parameter matrix, a 3×4 matrix consisting of the rotation matrix R (3×3) and the displacement vector T (3×1), is used to convert world coordinates to the camera coordinate system. K represents the camera's intrinsic parameter matrix (3×3), which describes the imaging system's focal length, principal point position, and other parameters and is used to map camera coordinates to the image pixel plane. (u,v) represents the final two-dimensional pixel position in the image, which is used to draw the center of the target's detection bounding box.

[0052] Precision association (motion trajectory matching): Single-target scenario: Using the Hungarian algorithm, the cost matrix contains: Position distance: Euclidean distance between the center of the ROI and the center of the detection box; appearance similarity: cosine distance of the feature vector extracted by ResNet-18.

[0053] Multi-target scenario: Build a graph neural network (GNN), input the target motion trajectory (UWB velocity vector) and appearance features (embedding vector extracted by ResNet-18), iteratively update the node state through three layers of message passing, output the probability matching matrix, and select matching pairs with a probability greater than 0.8.

[0054] Cost function: (λ is dynamically adjusted based on the time synchronization error, and the position weight is increased when the error is large).

[0055] 5. Determine whether the target position is in the overlapping area: According to the target location information Determine the current scene, specifically, determine whether the target is in the field of view overlapping monitoring area based on the current location information of the target and the monitoring mapping table.

[0056] 6. Calculate according to different scheduling scenarios If the current position is in a non-overlapping area, the scheduled camera is the camera monitoring the area, and the current camera is scheduled to track and output the video stream.

[0057] If the current location is in the overlapping area, the priority score of the surveillance camera in the overlapping area is calculated.

[0058] 7. Determine the camera to be scheduled based on the priority score. Specifically, schedule the camera with the highest priority score to track and output the video stream.

[0059] 8. Multi-source fusion, extended Kalman filtering, State vector: predict: Observation 1: UWB: Observation 2: Visual (only if confidence ≥ 0.3), Back projection obtains ; Occlusion processing: When the detection confidence is low or the bbox area drops by 50%, increase covariance, making filtering mainly dependent on UWB.

[0060] 9. Annotation generation and occlusion compensation: Output semantic annotations with absolute coordinates to ensure data continuity during occlusion.

[0061] Normal annotation mode: Fusion of visual detection frame and UWB calibration coordinates.

[0062] Visual loss processing: Predicted position based on UWB historical coordinates (10Hz) and motion model: Specifically, take the world coordinates after EKF update ,use Project back to the current camera pixel center, correct the bbox offset, generate a virtual box, and output structured annotations.

[0063] The three-dimensional coordinates of the UWB calibration of the present invention are integrated with visual semantics to output structured data including target category, absolute position (x, y, z), speed and time synchronization error, which can be directly used for high-precision spatial analysis.

[0064] The above specific implementation methods are detailed descriptions of the present invention. It cannot be considered that the specific implementation methods of the present invention are limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, they can make several simple deductions and substitutions without departing from the concept of the present invention, which should be regarded as falling within the scope of protection of the present invention.

Claims

1. A camera data annotation method based on UWB and target tracking, characterized in that: include: Using the absolute three-dimensional coordinates of the target provided by UWB, its coverage in the field of view of each camera is calculated in real time; Combined with motion trend prediction, camera priority scores are dynamically generated; The scheduling scenarios are divided into non-overlapping area scheduling and overlapping area scheduling based on the current field of view of the target. In the non-overlapping area scheduling scenario, the current camera is directly scheduled; in the overlapping area scheduling scenario, the camera with the highest score is selected for scheduling based on the camera priority score. When visual tracking fails, the UWB coordinates are directly used as spatial anchors to maintain trajectory continuity; when vision is available, the extended Kalman filter is used to integrate UWB positioning and visual observation into a smooth trajectory, and structured annotation data containing target identity, location, timestamp and appearance features are output in real time.

2. The camera data labeling method based on UWB and target tracking according to claim 1, characterized in that: In the overlapping area scheduling scenario, UWB coordinates and visual features are combined to perform cross-camera association.

3. The camera data labeling method based on UWB and target tracking according to claim 2, characterized in that: The cross-camera association of the combined UWB coordinates and visual features includes: First, the three-dimensional coordinates of the UWB tag are projected onto the image plane to generate a region of interest, and only the visual detection targets within the region of interest are associated; then, the Hungarian algorithm and graph neural network are combined to calculate the matching cost based on the target position, appearance features and motion consistency.

4. The camera data labeling method based on UWB and target tracking according to claim 3, characterized in that: The cross-camera association of the combined UWB coordinates and visual features further includes: Introduce a conflict resolution mechanism to prioritize targets with consistent movement directions.

5. The camera data labeling method based on UWB and target tracking according to claim 3 is characterized in that: The region of interest is calculated as follows: in, Indicates the absolute three-dimensional coordinates of the target, represents the external parameter matrix of the camera, K represents the internal parameter matrix of the camera, Represents the final two-dimensional pixel position on the image.

6. The camera data labeling method based on UWB and target tracking according to claim 3, characterized in that: Function of the matching cost: Among them, λ represents the weight.

7. The camera data labeling method based on UWB and target tracking according to claim 1, characterized in that: The absolute three-dimensional coordinates of the target provided by UWB include: In the offline phase, a calibration board with a UWB tag is deployed. By collecting multiple sets of coordinate corresponding points, the least squares method is used to solve the transformation matrix between the UWB global coordinate system and the camera coordinate system. During online operation, the calibration parameters are corrected in real time through the extended Kalman filter.

8. The camera data labeling method based on UWB and target tracking according to claim 1, characterized in that: The method of dynamically generating camera priority scores by combining motion trend prediction includes: Priority is calculated based on the target's distance from the camera and how far the target is expected to move within the camera's field of view.

9. The camera data labeling method based on UWB and target tracking according to claim 8, characterized in that: The calculation of the priority specifically includes: The method for calculating the priority score of each camera is: in, For camera The priority score, Target and camera The actual distance between Target on camera Estimated activity distance within the visual monitoring range; is the weight index, and the constraints are ; The weight of the weight index is determined by evaluating the quantitative index, and the method for determining the quantitative index is as follows: in, To evaluate quantitative indicators, and are the maximum and minimum priority scores corresponding to the cameras currently tracking the target.

10. The camera data labeling method based on UWB and target tracking according to claim 1, characterized in that: Set the detection confidence threshold. When the detection confidence is lower than the threshold, visual tracking is considered invalid and the system automatically switches to UWB-dominated tracking mode.

Images