A method of single-target tracking with trajectory prediction fusion

Abstract

The present application relates to the technical field of video single target tracking, and provides a single target tracking method fusing trajectory prediction, which comprises the following steps: acquiring a current video frame image and a last frame final target frame, using a basic tracker to perform parallel tracking on a target in the current video frame, and screening effective candidate target frames; constructing a motion state vector according to historical final target frames, predicting a current frame trajectory prediction frame, calculating the comprehensive reliability of the effective candidate frames, and obtaining reliable candidate frames; constructing a fusion feature vector, inputting the fusion feature vector into a meta-learning regression model, and obtaining a meta-learning prediction frame; calculating a drift risk value and judging a risk interval to which the drift risk value belongs, and selecting a corresponding scheme to output a current frame final target frame according to the risk interval. The present application can inhibit the influence of abnormal candidate frames on the final tracking result and subsequent state updating.

Description

Technical Field

[0001] This invention relates to the field of video single-target tracking technology, and more particularly to a method for single-target tracking that integrates trajectory prediction. Background Technology

[0002] Single-object tracking technology aims to continuously and accurately estimate the position and scale of a target in subsequent frames based on the target's position information in the first frame of a video. This technology has been widely applied in various fields such as intelligent surveillance, sports video analysis, human-computer interaction, robot vision, drone video analysis, and edge video processing. Existing single-object tracking methods can be mainly divided into correlation filtering methods, online learning methods, detection and tracking combined methods, and deep learning-based methods. Among them, traditional trackers, represented by CSRT, KCF, MIL, TLD, and MOSSE, have advantages such as simple deployment, low computational overhead, and strong interpretability. However, different trackers perform differently when facing complex situations such as occlusion, rapid motion, scale changes, pose changes, motion blur, and background interference. A single tracker often struggles to maintain stable and robust tracking performance in dynamically changing scenes.

[0003] To improve tracking robustness, existing research has proposed multi-tracker fusion schemes, typically employing simple averaging, fixed-weighted summaries, voting mechanisms, or selection strategies based on a single confidence level to output the final target box. However, most of these methods only focus on the reliability of individual candidate boxes in the current frame, failing to fully explore information such as the consistency among candidate boxes, the motion continuity between candidate boxes and historical trajectories, and the degree of divergence of the overall candidate set. When a base tracker drifts, its erroneous candidate boxes may still participate in the fusion process, thus contaminating the final output. If erroneous outputs are used for state updates in subsequent frames, drift errors will accumulate and propagate over time, further degrading tracking performance.

[0004] Furthermore, existing trajectory prediction methods, such as those based on Kalman filtering, are typically used only as short-term compensation measures or as independent trajectory outputs, failing to form an organic linkage with the multi-tracker fusion process. Specifically, the trajectory prediction results lack a unified constraint and closed-loop coordination mechanism among candidate box quality assessment, fusion feature construction, drift risk judgment, and state update suppression. Therefore, when facing challenging scenarios such as short-term occlusion, similar target intersections, rapid movement and turning, and background similarity interference, existing methods are still prone to problems such as target box jumps, target drift, inaccurate scale estimation, or tracking failures, and their overall robustness and adaptability still need to be improved. Summary of the Invention

[0005] To address the above issues, this invention constructs a closed-loop tracking mechanism consisting of multi-basic tracker candidate generation, trajectory prior prediction, three-dimensional reliability evaluation, structured meta-learning fusion, drift risk quantification, dynamic weight correction, and controlled state update. This mechanism ensures that the trajectory prediction result is not merely a replacement output after tracking failure, but rather serves as a multi-stage constraint signal throughout candidate box evaluation, fusion prediction, drift determination, dynamic correction, and state update, thereby suppressing the impact of abnormal candidate boxes on the final tracking result and subsequent state updates.

[0006] According to an embodiment of the present invention, a method for single-target tracking based on fused trajectory prediction is provided.

[0007] In a first aspect of the invention, a method for single-target tracking based on fused trajectory prediction is provided. The method includes: Step S01: Obtain the current video frame image and the final target box of the previous frame, and use at least two basic trackers to track the target in the current video frame in parallel to filter valid candidate target boxes; Step S02: Construct motion state vectors based on historical final target boxes, obtain the trajectory prediction box of the current frame through trajectory prediction model, and calculate the comprehensive credibility of the effective candidate boxes and the reliability-weighted candidate boxes by combining the effective candidate boxes and the final target boxes of the previous frame. Step S03: Construct a fusion feature vector based on the effective candidate target boxes, comprehensive confidence, trajectory prediction boxes, the final target box of the previous frame, and the candidate set statistics, and input the fusion feature vector into the meta-learning regression model to obtain the meta-learning prediction boxes; Step S04: Calculate the drift risk value based on the meta-learning prediction box, trajectory prediction box, reliability-weighted candidate box, and the degree of divergence of the candidate set, and determine the risk interval to which it belongs. Select the corresponding scheme based on the risk interval to output the final target box of the current frame.

[0008] Furthermore, the basic tracker in step S01 includes at least two of CSRT, KCF, MIL, TLD, MOSSE, and BOOSTING.

[0009] Furthermore, the selection method for valid candidate target boxes in step S01 is as follows: the basic tracker returns a success flag, the candidate box does not cross the boundary, and the area is not less than the minimum threshold or the area is not greater than a preset ratio of the image area.

[0010] Furthermore, the motion state vector in step S02 includes the target center point coordinates, the target center point velocity, the target bounding box width, and the target bounding box height.

[0011] Furthermore, the specific steps for calculating the comprehensive credibility of valid candidate boxes and the reliability-weighted candidate boxes in step S02 are as follows: Calculate the motion continuity score, trajectory consistency score, and multi-tracker group consistency score for each valid candidate box: , , , in, and The first Frame number The and the first One valid candidate target box, and For the first The x and y coordinates of the center point of each candidate box and The first The x and y coordinates of the center point of the final target bounding box of the frame For trajectory prediction bounding boxes, , These are the image width and height, respectively. For the continuity of motion fraction, The trajectory consistency score. For group consistency score, , Mark the candidate box as valid. This is the normalization coefficient for the movement distance. It is the intersection-union ratio function; Calculate the overall credibility, normalized weights, and reliability-weighted candidate boxes: ,and , , , in, To assess overall credibility, To normalize the weights, For reliability-weighted candidate boxes, Temperature coefficient normalized by softmax These are the weighting coefficients.

[0012] Furthermore, the meta-learning regression model described in step S03 is one of random forest regression, gradient boosting tree, XGBoost, ridge regression, or lightweight multilayer perceptron.

[0013] Furthermore, the formula for calculating the drift risk value mentioned in step S04 is as follows: , in, These are the weighting coefficients. , To normalize the center displacement, To determine the degree of divergence in the candidate set, The degree of scale variation anomaly: , , , In the formula, For the first The x-coordinate of the center point of the frame learning prediction box. For the first The ordinate of the center point of the frame-learned prediction box. The x-coordinate of the center point of the final target bounding box in the previous frame. The ordinate of the center point of the final bounding box in the previous frame. For the first The width of the frame-based learning prediction box. The width of the final target bounding box in the previous frame. For the first The height of the prediction box in the frame learning. The height of the final target bounding box in the previous frame. This represents the number of valid candidate bounding boxes in the current frame. and The first Frame number The and the first One valid candidate target box.

[0014] Furthermore, the step S04, which involves selecting the corresponding solution based on the risk range and outputting the final target box of the current frame, specifically involves: like If the drift risk is low, the meta-learning prediction box or the fusion box with the meta-learning prediction box as the main component is output. like If the current frame is determined to have a medium drift risk, dynamic correction is triggered, and the corresponding solution is selected based on the risk range to output the final target bounding box of the current frame. : , in, These are the fusion weights for the meta-learning predicted bounding box, the trajectory predicted bounding box, and the reliability-weighted candidate bounding box, respectively. ; like If the current frame is deemed to have a high risk of drift, the meta-learning prediction box will be reduced. The fusion weights improve the trajectory prediction box. and reliability-weighted candidate boxes The fusion weights, or the trajectory prediction box. With reliability-weighted candidate boxes The fusion result is used as the final target box of the current frame. ;in, and To preset the drift risk threshold, and ;when When it increases, Decrease and Increase.

[0015] Furthermore, it also includes step S05: based on the risk range to which the drift risk value belongs, perform corresponding controlled state updates on the base tracker and trajectory prediction model, if... If so, the base tracker and trajectory prediction model will be updated normally; if Then, the trajectory prediction model is updated using the corrected final target box, and a low-learning-rate update or selective update is performed on the base tracker. The update rate can be expressed as... ;like If the update is frozen, only the reliable state or trajectory prediction state of the previous frame is retained.

[0016] In a second aspect of the invention, an apparatus for single-target tracking based on fused trajectory prediction is provided. The apparatus includes: Effective candidate target box filtering module: used to acquire the current video frame image and the final target box of the previous frame, and use at least two basic trackers to track the target in the current video frame in parallel to filter effective candidate target boxes; The reliability-weighted candidate box calculation module is used to construct motion state vectors based on historical final target boxes, obtain the trajectory prediction box of the current frame through the trajectory prediction model, and calculate the comprehensive credibility of the effective candidate boxes and the reliability-weighted candidate boxes by combining the effective candidate boxes and the final target boxes of the previous frame. Meta-learning prediction box acquisition module: It is used to construct a fused feature vector based on the effective candidate target boxes, comprehensive confidence, trajectory prediction boxes, the final target box of the previous frame and the candidate set statistics, and input the fused feature vector into the meta-learning regression model to obtain the meta-learning prediction box; The current frame final target box acquisition module is used to calculate the drift risk value and determine its risk interval based on the meta-learning prediction box, trajectory prediction box, reliability weighted candidate box and the degree of divergence of the candidate set. Based on the risk interval, the module selects the corresponding scheme and outputs the current frame final target box. Controlled State Update Module: This module is used to perform corresponding controlled state updates on the base tracker and trajectory prediction model based on the risk range to which the drift risk value belongs.

[0017] This invention constructs a closed-loop tracking mechanism consisting of multi-basic tracker candidate generation, trajectory prior prediction, three-dimensional credibility evaluation, structured meta-learning fusion, drift risk quantification, dynamic weight correction, and controlled state update. This mechanism ensures that the trajectory prediction result is not merely a substitute output after tracking failure, but rather serves as a multi-stage constraint signal that runs through candidate box evaluation, fusion prediction, drift determination, dynamic correction, and state update. This suppresses the impact of abnormal candidate boxes on the final tracking result and subsequent state update.

[0018] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0019] The beneficial effects of this invention are: 1. By using three-dimensional credibility evaluation, abnormal candidate boxes are suppressed before they enter the fusion model, reducing the impact of drift of a single base tracker on the final result; 2. By using the trajectory prediction box as a multi-stage constraint signal, the historical motion prior is used not only for short-term compensation, but also for constraining candidate quality, fusion input, and drift risk. 3. By learning features through structured meta-learning, the fusion model can simultaneously learn the relationships between visual candidates, motion priors, historical states, and candidate set divergences. 4. Through dynamic correction triggered by drift risk, the meta-learning prediction box maintains its dominant role in normal scenarios, while reducing its weight in determining the final output when there is occlusion, rapid change of direction, or intersection of similar targets. This avoids directly outputting abnormal meta-learning prediction boxes and improves stability in complex scenarios. 5. By controlling state updates, the probability of abnormal results entering the state update link of subsequent frames is reduced, thus suppressing the cumulative propagation of drift in the time series. 6. It adopts a traditional basic tracker, lightweight trajectory prediction and lightweight regression model, which has good interpretability and engineering deployment value, and is suitable for target tracking at low computing power edge. Attached Figure Description

[0020] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein: Figure 1 A flowchart of a single-target tracking method based on fused trajectory prediction according to an embodiment of the present invention is shown; Figure 2 A block diagram of the multi-tracker candidate generation and meta-learning fusion module according to Embodiment 1 of the present invention is shown; Figure 3A schematic diagram of dynamic weight adjustment triggered by drift risk according to Embodiment 1 of the present invention is shown; Figure 4 A schematic diagram of the controlled state update mechanism according to Embodiment 1 of the present invention is shown; Figure 5 A screenshot of the target tracking simulation of the fusion output according to Embodiment 2 of the present invention is shown; Figure 6 A comparative schematic diagram of meta-learning prediction boxes, trajectory prediction boxes, reliability-weighted candidate boxes, and final correction boxes according to an embodiment of the present invention is shown; Figure 7 The figure shows a comparison curve of the center position error of each comparison algorithm according to an embodiment of the present invention on a Basketball test sequence; Figure 8 The accuracy and success rate curves of various comparison algorithms according to embodiments of the present invention on basketball test sequences are shown. Figure 9 A bar chart comparing the comprehensive quantitative indicators of various comparison algorithms according to embodiments of the present invention on a Basketball test sequence is shown. Figure 10 A block diagram of an apparatus for single-target tracking based on fusion trajectory prediction according to an embodiment of the present invention is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] According to an embodiment of the present invention, a single-target tracking method based on fusion trajectory prediction is proposed. By constructing a closed-loop tracking mechanism consisting of multi-base tracker candidate generation, trajectory prior prediction, three-dimensional credibility evaluation, structured meta-learning fusion, drift risk quantification, dynamic weight correction, and controlled state update, the trajectory prediction result is no longer just a substitute output after tracking failure, but serves as a multi-stage constraint signal that runs through candidate box evaluation, fusion prediction, drift determination, dynamic correction, and state update, thereby suppressing the influence of abnormal candidate boxes on the final tracking result and subsequent state update.

[0023] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0024] Figure 1This is a schematic flowchart of a single-target tracking method based on fused trajectory prediction according to an embodiment of the present invention. The method includes: Step S01: Obtain the current video frame image and the final target box of the previous frame, and use at least two basic trackers to track the target in the current video frame in parallel to filter valid candidate target boxes; Step S02: Construct motion state vectors based on historical final target boxes, obtain the trajectory prediction box of the current frame through trajectory prediction model, and calculate the comprehensive credibility of the effective candidate boxes and the reliability-weighted candidate boxes by combining the effective candidate boxes and the final target boxes of the previous frame. Step S03: Construct a fusion feature vector based on the effective candidate target boxes, comprehensive confidence, trajectory prediction boxes, the final target box of the previous frame, and the candidate set statistics, and input the fusion feature vector into the meta-learning regression model to obtain the meta-learning prediction boxes; Step S04: Calculate the drift risk value based on the meta-learning prediction box, trajectory prediction box, reliability weighted candidate box and the degree of divergence of the candidate set, and determine the risk interval to which it belongs. Select the corresponding scheme according to the risk interval to output the final target box of the current frame. Step S05: Based on the risk range to which the drift risk value belongs, perform the corresponding controlled state update for the base tracker and trajectory prediction model.

[0025] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0026] To provide a clearer explanation of the above-described method for single-target tracking based on fused trajectory prediction, three specific embodiments are described below. However, it is worth noting that these embodiments are only for better illustrating the present invention and do not constitute an undue limitation on the present invention.

[0027] The following three specific examples will further illustrate the single-target tracking method based on fused trajectory prediction in more detail.

[0028] Example 1: Closed-loop single-target tracking method.

[0029] Step S01: Obtain the current video frame image and the final target bounding box of the previous frame, and use at least two basic trackers to perform parallel tracking of the target in the current video frame to filter valid candidate target bounding boxes.

[0030] Initialize at least two base trackers and trajectory prediction models. The base trackers can be selected from at least two of CSRT, KCF, MIL, TLD, MOSSE, and BOOSTING.

[0031] In the current frame processing, the system obtains the first... The image frame and the final bounding box from the previous frame. Multiple base trackers output candidate bounding boxes respectively. ,in, and For the first Coordinates of the center point of each candidate box and For the first The width and height of each candidate box. , The number of base trackers.

[0032] When a base tracker returns a failure flag, a candidate box goes out of bounds, has an area smaller than a minimum threshold, or has an area larger than a preset proportion of the image area, the candidate box is marked as invalid and then marked as valid. =0; otherwise let =1, filtering out valid candidate bounding boxes. For any original candidate bounding box that meets the validity condition, it is recorded as a valid candidate bounding box. This forms a set of valid candidate target boxes for the current frame.

[0033] If the current frame is the first frame, the target bounding box of the first frame, either manually labeled or provided by the target detection module, will be used as the final target bounding box of the current frame, and the target bounding box of the first frame will be used to initialize each basic tracker and trajectory prediction model.

[0034] Step S02: Construct motion state vectors based on historical final target boxes, obtain the trajectory prediction box of the current frame through trajectory prediction model, and calculate the comprehensive credibility of the effective candidate boxes and the reliability-weighted candidate boxes by combining the effective candidate boxes and the final target boxes of the previous frame.

[0035] According to recent Frame final target bounding box constructs motion state vector , The historical window length is preferably between 3 and 10. The motion state vector includes the target center point coordinates, target center point velocity, target bounding box width, and target bounding box height, and can be represented as: ,in, and The first The x and y coordinates of the center point of the final target bounding box in the frame. and The first The width and height of the final target bounding box of the frame. and Based on the most recent The final target box estimation of the frame yields the average lateral velocity and the average longitudinal velocity. In one implementation, , ,in, and The first The x and y coordinates of the center point of the final target bounding box of the frame.

[0036] Trajectory prediction boxes are obtained using Kalman prediction models, linear motion models, or lightweight temporal prediction models. The trajectory prediction bounding box does not directly replace the visual tracking results, but is used for subsequent credibility evaluation, feature fusion construction, drift risk assessment, and dynamic correction.

[0037] For each valid candidate box, three scores are calculated: motion continuity score, trajectory consistency score, and multi-tracker group consistency score. Let the... Frame number The number of valid candidate target boxes is The final target bounding box in the previous frame was The trajectory prediction box is The image width and height are respectively , The motion continuity score, trajectory consistency score, and group consistency score can be determined by the following formulas: , , , in, For the continuity of motion fraction, The trajectory consistency score. For group consistency score, This represents the number of valid candidate bounding boxes in the current frame. , Mark the candidate box as valid. This is the normalization coefficient for the movement distance. Let be the intersection-union ratio function. This is used to measure the degree of overlap between two bounding boxes.

[0038] The overall credibility of the candidate boxes is obtained by weighting the motion continuity score, trajectory consistency score, and group consistency score. All are greater than or equal to 0, and ; The temperature coefficient is normalized using softmax. The candidate boxes weighted by combining confidence, normalization weight, and reliability can be represented as follows: ,and , , , in, To assess overall credibility, To normalize the weights, Reliability-weighted candidate boxes.

[0039] Step S03: Construct a fusion feature vector based on the effective candidate target boxes, comprehensive confidence, trajectory prediction boxes, the final target box of the previous frame, and the candidate set statistics, and input the fusion feature vector into the meta-learning regression model to obtain the meta-learning prediction boxes.

[0040] Constructing fused feature vectors The fused feature vector includes effective candidate bounding boxes, overall confidence score, trajectory prediction bounding boxes, the final bounding box from the previous frame, and statistical features of the effective candidate set. Statistical features of the effective candidate set include at least one of the following: number of effective candidate bounding boxes, mean center point value of effective candidate bounding boxes, mean width and height value of effective candidate bounding boxes, variance of center point value of effective candidate bounding boxes, variance of width and height value of effective candidate bounding boxes, maximum overall confidence score, minimum overall confidence score, average overall confidence score, and maximum intersection-union ratio (IUU) between effective candidate bounding boxes and trajectory prediction bounding boxes.

[0041] The fused feature vector can be represented as ,in, To fuse feature vectors, For the first Frame number One valid candidate target box, For the first Frame number The overall credibility of the valid candidate target boxes. The current frame trajectory prediction box, The final target bounding box of the previous frame, where, The statistical features of the effective candidate set include at least one of the following: the number of effective candidate boxes, the mean of candidate boxes, the variance of candidate boxes, the average intersection-union ratio (IUU) among candidate boxes, and the confidence statistic of candidate boxes. When the number of effective candidate boxes is inconsistent in different video frames, zero-padded, count truncated, or validity masked processing is applied to the candidate boxes and their corresponding confidence scores to improve the fused feature vector. It meets the fixed input dimension requirement of the meta-learning regression model.

[0042] like Figure 2 As shown, the feature vectors will be fused. Input the trained meta-learning regression model Obtain the meta-learning prediction box Meta-learning regression models can employ random forest regression, gradient boosting trees, XGBoost, ridge regression, or lightweight multilayer perceptrons.

[0043] Specifically, the training method for the meta-learning regression model is as follows: During the training phase, video sequences containing manually labeled target boxes are selected. Multiple basic trackers and trajectory prediction models are run on each frame to calculate candidate box confidence and candidate set statistics, forming training samples. The coordinates of the center point and the width and height of the real target boxes are used as supervision labels to train the meta-learning regression model. After training, the model parameters are fixed during the inference phase, and meta-learning prediction boxes are output only based on the candidate boxes, confidence, and trajectory priors in the current video frame. To improve generalization ability, the training samples preferably cover scenarios such as no occlusion, short-term occlusion, intersection of similar targets, rapid direction changes, scale changes, and background interference, enabling the model to learn the correspondence between candidate set statistics and real target boxes under different risk states.

[0044] Step S04: Calculate the drift risk value based on the meta-learning prediction box, trajectory prediction box, reliability-weighted candidate box, and the degree of divergence of the candidate set, and determine the risk interval to which it belongs. Select the corresponding scheme based on the risk interval to output the final target box of the current frame.

[0045] Calculate drift risk value The drift risk value can be obtained by weighting at least two of the following: the intersection-union ratio (IU) loss between the meta-learned predicted box and the trajectory predicted box; the normalized center displacement of the meta-learned predicted box relative to the final target box in the previous frame; the degree of divergence among multiple candidate target boxes; and the degree of scale change anomaly of the meta-learned predicted box relative to the final target box in the previous frame. For example: , in, , To normalize the center displacement, To determine the degree of divergence in the candidate set, This represents the degree of scale variation anomaly.

[0046] , , , In the formula, For the first The x-coordinate of the center point of the frame learning prediction box. For the first The ordinate of the center point of the frame-learned prediction box. The x-coordinate of the center point of the final target bounding box in the previous frame. The ordinate of the center point of the final bounding box in the previous frame. For the first The width of the frame-based learning prediction box. The width of the final target bounding box in the previous frame. For the first The height of the prediction box in the frame learning. The height of the final target bounding box in the previous frame. This represents the number of valid candidate bounding boxes in the current frame. and The first Frame number The and the first One valid candidate target box.

[0047] Disagreement of valid candidate set This is used to indicate the degree of positional difference between multiple valid candidate bounding boxes. When the overlap between multiple valid candidate bounding boxes is high, Smaller; when there is a large discrepancy between multiple valid candidate bounding boxes. Larger. When hour, It can be set to a preset high-risk value, or determined based on the degree of deviation between the valid candidate target boxes and the trajectory prediction boxes.

[0048] like If the current frame is determined to have a low risk of drift, then the meta-learned prediction box is output. As the final target box of the current frame Or output a meta-learning prediction box. The primary fusion box is used as the final target box in the current frame. .

[0049] like If the current frame is determined to have a medium risk of drift, dynamic correction is triggered, and the prediction box is based on meta-learning. Trajectory prediction box and reliability-weighted candidate boxes Perform weighted fusion to obtain the final target bounding box of the current frame. : , Among them, among them, These are reliability-weighted candidate boxes obtained by weighting the overall credibility of each valid candidate target box. , These are the fusion weights for the meta-learning predicted bounding box, the trajectory predicted bounding box, and the reliability-weighted candidate bounding box, respectively.

[0050] Final target box in current frame The output is used to update the historical state of the next frame.

[0051] like Figure 3 As shown, if If the current frame is deemed to have a high risk of drift, the meta-learning prediction box will be reduced. The fusion weights improve the trajectory prediction box. and reliability-weighted candidate boxes The fusion weights, or the trajectory prediction box. With reliability-weighted candidate boxes The fusion result is used as the final target box of the current frame. ;in, and To preset the drift risk threshold, and ;when When it increases, Decrease and Increase.

[0052] Step S05: Based on the risk range to which the drift risk value belongs, perform the corresponding controlled state update for the base tracker and trajectory prediction model.

[0053] Specifically, such as Figure 4 As shown, if If so, the base tracker and trajectory prediction model will be updated normally; if Then, the trajectory prediction model is updated using the corrected final target box, and a low-learning-rate update or selective update is performed on the base tracker. The update rate can be expressed as... ;like If the update is frozen, only the reliable state or trajectory prediction state of the previous frame is retained to avoid erroneous visual features contaminating subsequent frames.

[0054] Example 2: Application scenarios for sports videos.

[0055] like Figures 5-6 As shown, taking a basketball game broadcast video as an example, the video sequence resolution is 1920×1080, and the target is a player in a blue jersey carrying the ball. The bounding box of the first frame image, provided manually or by the detection module, is... Where (612, 286) are the coordinates of the center point of the target bounding box, and 96 and 214 are the width and height of the target bounding box, respectively. The system initializes the three basic trackers CSRT, KCF and MOSSE with this first frame target bounding box, and initializes the Kalman trajectory prediction model.

[0056] In frame 18, the target athlete moves quickly to the right and forward, and the three basic trackers output candidate boxes respectively. =(648,296,100,216) =(651,295,98,214) =(702,310,96,210), trajectory prediction box =(645,294,99,216). Based on the aforementioned formula, the overall confidence levels of the three candidate boxes are calculated to be 0.91, 0.87, and 0.42, respectively, with softmax normalized weights of 0.44, 0.41, and 0.15. The reliability-weighted candidate box is... =(650,296,99,215). This shows that the third candidate box, which deviates from the motion trajectory and is inconsistent with the group of candidates, has its weight reduced.

[0057] In frame 36, another athlete wearing a similar jersey briefly cross-occludes the target, causing the candidate boxes output by KCF to shift towards the interfering target. However, the CSRT candidate boxes, MOSSE candidate boxes, and trajectory prediction boxes remain the same. The focus remains on the vicinity of the target athlete. The system calculates the meta-learning prediction box. With trajectory prediction box The crossover ratio loss is 0.18, and the candidate divergence level is... 6. Degree of anomalous scale changes Overall drift risk value Below the first threshold Therefore, the output meta-learned prediction box is... As the final target box .

[0058] In frame 52, the target is approximately 60% obscured by the defender. Both the KCF and MOSSE candidate boxes shift towards the obscured target, indicating a significant divergence among the candidates. The meta-learning prediction box rose to 0.48. With trajectory prediction box The crossover loss ratio increased to 0.41, and the overall drift risk value... Above the first threshold And below the second threshold The system triggers dynamic correction, adjusting the fusion weights of the meta-learning predicted bounding box, trajectory predicted bounding box, and reliability-weighted candidate bounding box from the normal values ​​of 0.60, 0.20, and 0.20 to 0.30, 0.45, and 0.25, thus obtaining the final target bounding box for the current frame. Furthermore, a low learning rate is used to update the base tracker to avoid occluding the appearance features of the target and contaminating subsequent frames.

[0059] like Figure 7As shown, in the test of the same video segment, using KCF alone resulted in two target switchings with an average center error of 18.6 pixels; using CSRT alone did not result in significant target switching but had a higher average processing time with an average center error of 12.4 pixels; using only Kalman trajectory prediction resulted in continuous deviations after sudden stops and turns with an average center error of 21.3 pixels. This invention, by fusing trajectory prediction, candidate confidence evaluation, and dynamic correction, significantly reduces the number of target switchings, lowers the average center error to 8.7 pixels, and allows the target bounding box to re-fit the target athlete in most occlusion recovery frames. Therefore, compared to a single visual tracker or a single trajectory prediction method, it has better resistance to occlusion, resistance to interference from similar targets, and rapid motion adaptation capabilities.

[0060] like Figure 8 As shown, the accuracy curve and success rate curve are used to characterize the performance differences of different tracking methods in terms of target center localization accuracy and target bounding box overlap matching degree. Specifically, the accuracy curve reflects the proportion of frames where the predicted target bounding box center position error is less than different thresholds, and the success rate curve reflects the proportion of frames where the intersection-union ratio (IUU) between the predicted target bounding box and the ground truth target bounding box is greater than different thresholds. Figure 8 It can be seen that the method of the present invention can maintain high accuracy and success rate under multiple threshold conditions, indicating that it has good target localization ability and scale estimation stability in complex video scenarios.

[0061] like Figure 9 As shown, the comprehensive quantitative indicators include the area under the success rate curve, 20-pixel accuracy, and average error. A larger area under the success rate curve, higher 20-pixel accuracy, and lower average error indicate better overall performance of the tracking method. Figure 9 It can be seen that the method of the present invention has better overall tracking performance than a single basic tracker and conventional fusion methods. It can effectively reduce the impact of target drift and candidate box anomalies, thereby improving the accuracy and robustness of single target tracking.

[0062] Based on the same inventive concept, this invention also proposes a device for single-target tracking that integrates trajectory prediction. The implementation of this device can be found in the implementation of the method described above; repeated details will not be repeated. Figure 10 As shown, the device 100 includes: Effective candidate target box filtering module 101: used to acquire the current video frame image and the final target box of the previous frame, and use at least two basic trackers to perform parallel tracking of the target in the current video frame to filter effective candidate target boxes; Reliability-weighted candidate box calculation module 102: It is used to construct motion state vectors based on historical final target boxes, obtain the trajectory prediction box of the current frame through trajectory prediction model, and calculate the comprehensive credibility of the effective candidate boxes and the reliability-weighted candidate boxes by combining the effective candidate boxes and the final target boxes of the previous frame. Meta-learning prediction box acquisition module 103: is used to construct a fusion feature vector based on the effective candidate target boxes, comprehensive confidence, trajectory prediction boxes, the final target box of the previous frame and the candidate set statistics, and input the fusion feature vector into the meta-learning regression model to obtain the meta-learning prediction box; The current frame final target box acquisition module 104 is used to calculate the drift risk value and determine its risk interval based on the meta-learning prediction box, trajectory prediction box, reliability weighted candidate box and the degree of divergence of the candidate set, and select the corresponding scheme to output the current frame final target box according to the risk interval. Controlled State Update Module 105: Used to perform corresponding controlled state updates on the base tracker and trajectory prediction model based on the risk range to which the drift risk value belongs.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0064] Furthermore, although the operations are described in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0065] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for single-target tracking that integrates trajectory prediction, characterized in that, The method includes: Step S01: Obtain the current video frame image and the final target box of the previous frame, and use at least two basic trackers to track the target in the current video frame in parallel to filter valid candidate target boxes; Step S02: Construct motion state vectors based on historical final target boxes, obtain the trajectory prediction box of the current frame through trajectory prediction model, and calculate the comprehensive credibility of the effective candidate boxes and the reliability-weighted candidate boxes by combining the effective candidate boxes and the final target boxes of the previous frame. The specific steps for calculating the comprehensive credibility of valid candidate boxes and the reliability-weighted candidate boxes are as follows: Calculate the motion continuity score, trajectory consistency score, and multi-tracker group consistency score for each valid candidate box: ， ， ， in, and The first Frame number The and the first One valid candidate target box, and For the first The x and y coordinates of the center point of each candidate box and The first The x and y coordinates of the center point of the final target bounding box in the frame. For trajectory prediction bounding boxes, , These are the image width and height, respectively. For the continuity of motion fraction, The trajectory consistency score. For group consistency score, , Mark the candidate box as valid. This is the normalization coefficient for the movement distance. It is the intersection-union ratio function; Calculate the overall credibility, normalized weights, and reliability-weighted candidate boxes: ,and , ， ， in, To assess overall credibility, To normalize the weights, For reliability-weighted candidate boxes, Temperature coefficient normalized by softmax These are the weighting coefficients; Step S03: Construct a fusion feature vector based on the effective candidate target boxes, comprehensive confidence, trajectory prediction boxes, the final target box of the previous frame, and the candidate set statistics, and input the fusion feature vector into the meta-learning regression model to obtain the meta-learning prediction boxes; Step S04: Calculate the drift risk value based on the meta-learning prediction box, trajectory prediction box, reliability weighted candidate box, and the degree of divergence of the candidate set, and determine the risk interval to which it belongs. Select the corresponding scheme according to the risk interval to output the final target box of the current frame. The formula for calculating the drift risk value is as follows: ， in, These are the weighting coefficients. , To normalize the center displacement, To determine the degree of divergence in the candidate set, The degree of scale variation anomaly, Meta-learning prediction boxes; ， ， ， In the formula, For the first The x-coordinate of the center point of the frame learning prediction box. For the first The ordinate of the center point of the frame-learned prediction box. The x-coordinate of the center point of the final target bounding box in the previous frame. The ordinate of the center point of the final bounding box in the previous frame. For the first The width of the frame-based learning prediction box. The width of the final target bounding box in the previous frame. For the first The height of the prediction box in the frame learning. The height of the final target bounding box in the previous frame. This represents the number of valid candidate bounding boxes in the current frame. and The first Frame number The and the first One valid candidate target box.

2. The method for single-target tracking based on fusion trajectory prediction according to claim 1, characterized in that, The basic tracker mentioned in step S01 includes at least two of CSRT, KCF, MIL, TLD, MOSSE, and BOOSTING.

3. The method for single-target tracking based on trajectory prediction according to claim 1, characterized in that, The method for selecting effective candidate target boxes in step S01 is as follows: the basic tracker returns a success flag, the candidate box does not cross the boundary, and the area is not less than the minimum threshold or the area is not greater than a preset ratio of the image area.

4. The method for single-target tracking based on fused trajectory prediction according to claim 1, characterized in that, The motion state vector in step S02 includes the target center point coordinates, the target center point velocity, the target bounding box width, and the target bounding box height.

5. The method for single-target tracking based on fused trajectory prediction according to claim 1, characterized in that, The meta-learning regression model mentioned in step S03 is one of random forest regression, gradient boosting tree, XGBoost, ridge regression, or lightweight multilayer perceptron.

6. The method for single-target tracking based on fused trajectory prediction according to claim 1, characterized in that, The step S04, which involves selecting the corresponding solution based on the risk range and outputting the final target bounding box of the current frame, specifically involves: like If the drift risk is low, the meta-learning prediction box or the fusion box with the meta-learning prediction box as the main component is output. like If the current frame is determined to have a medium drift risk, dynamic correction is triggered, and the corresponding solution is selected based on the risk range to output the final target bounding box of the current frame. : ， in, These are the fusion weights for the meta-learning predicted bounding box, the trajectory predicted bounding box, and the reliability-weighted candidate bounding box, respectively. ; like If the current frame is deemed to have a high risk of drift, the meta-learning prediction box will be reduced. The fusion weights improve the trajectory prediction box. and reliability-weighted candidate boxes The fusion weights, or the trajectory prediction box. With reliability-weighted candidate boxes The fusion result is used as the final target box of the current frame. ;in, and To preset the drift risk threshold, and ;when When it increases, Decrease and Increase.

7. The method for single-target tracking based on fused trajectory prediction according to claim 6, characterized in that, It also includes step S05: based on the risk range to which the drift risk value belongs, perform corresponding controlled state updates on the base tracker and trajectory prediction model, if If so, the base tracker and trajectory prediction model will be updated normally; if Then, the trajectory prediction model is updated using the corrected final target box, and a low-learning-rate update or selective update is performed on the base tracker. The update rate can be expressed as... ;like If the update is frozen, only the reliable state or trajectory prediction state of the previous frame is retained.

8. A device for single-target tracking that integrates trajectory prediction, characterized in that, The device implements the method as described in any one of claims 1 to 7, comprising: Effective candidate target box filtering module: used to acquire the current video frame image and the final target box of the previous frame, and use at least two basic trackers to track the target in the current video frame in parallel to filter effective candidate target boxes; The reliability-weighted candidate box calculation module is used to construct motion state vectors based on historical final target boxes, obtain the trajectory prediction box of the current frame through the trajectory prediction model, and calculate the comprehensive credibility of the effective candidate boxes and the reliability-weighted candidate boxes by combining the effective candidate boxes and the final target boxes of the previous frame. The specific steps for calculating the comprehensive credibility of valid candidate boxes and the reliability-weighted candidate boxes are as follows: Calculate the motion continuity score, trajectory consistency score, and multi-tracker group consistency score for each valid candidate box: ， ， ， in, and The first Frame number The and the first One valid candidate target box, and For the first The x and y coordinates of the center point of each candidate box and The first The x and y coordinates of the center point of the final target bounding box in the frame. For trajectory prediction bounding boxes, , These are the image width and height, respectively. For the continuity of motion fraction, The trajectory consistency score. For group consistency score, , Mark the candidate box as valid. This is the normalization coefficient for the movement distance. It is the intersection-union ratio function; Calculate the overall credibility, normalized weights, and reliability-weighted candidate boxes: ,and , ， ， in, To assess overall credibility, To normalize the weights, For reliability-weighted candidate boxes, Temperature coefficient normalized by softmax These are the weighting coefficients; Meta-learning prediction box acquisition module: It is used to construct a fused feature vector based on the effective candidate target boxes, comprehensive confidence, trajectory prediction boxes, the final target box of the previous frame and the candidate set statistics, and input the fused feature vector into the meta-learning regression model to obtain the meta-learning prediction box; The current frame final target box acquisition module is used to calculate the drift risk value and determine its risk interval based on the meta-learning prediction box, trajectory prediction box, reliability weighted candidate box and the degree of divergence of the candidate set. Based on the risk interval, the module selects the corresponding scheme and outputs the current frame final target box. Controlled State Update Module: This module is used to perform corresponding controlled state updates on the base tracker and trajectory prediction model based on the risk range to which the drift risk value belongs. The formula for calculating the drift risk value is as follows: ， in, These are the weighting coefficients. , To normalize the center displacement, To determine the degree of divergence in the candidate set, The degree of scale variation anomaly, Meta-learning prediction boxes; ， ， ， In the formula, For the first The x-coordinate of the center point of the frame learning prediction box. For the first The ordinate of the center point of the frame-learned prediction box. The x-coordinate of the center point of the final target bounding box in the previous frame. The ordinate of the center point of the final bounding box in the previous frame. For the first The width of the frame-based learning prediction box. The width of the final target bounding box in the previous frame. For the first The height of the prediction box in the frame learning. The height of the final target bounding box in the previous frame. This represents the number of valid candidate bounding boxes in the current frame. and The first Frame number The and the first One valid candidate target box.

Images