Single-target tracking target error retention elimination method and judgment system thereof
By employing a multi-dimensional feature complementarity verification method, the problem of false delay in single-target tracking algorithms under complex scenarios is solved, improving the accuracy and adaptability of target loss judgment and achieving a stable false delay elimination effect.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JINDIE SPACETIME (BEIJING) TECHNOLOGY CO LTD
- Filing Date
- 2026-01-20
- Publication Date
- 2026-05-19
AI Technical Summary
Existing single-target tracking algorithms are prone to tracking errors when faced with situations such as target occlusion, rapid motion blur, or target moving out of the field of view. Furthermore, existing confidence thresholding methods have poor adaptability and insufficient robustness in different scenarios, failing to meet the needs of complex scenarios.
A multi-dimensional feature complementary verification method is adopted, including edge logic judgment, target classification judgment, and motion constraint judgment. By setting confidence thresholds for different regions, using a classification model to identify target categories, and using Kalman filtering to predict motion trajectories, a comprehensive judgment is made on whether the target is lost.
It improves the accuracy and scene adaptability of target loss judgment, achieves efficient and stable single target tracking error elimination effect, and reduces the probability of misjudgment and drift.
Smart Images
Figure CN122066731A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision technology, and in particular relates to a method for eliminating false delays in single-target tracking and its judgment system. Background Technology
[0002] Single-object tracking algorithms refer to techniques that, given the initial position or feature description of a single target in the first frame of a continuous video sequence, continuously and accurately locate the target in subsequent frames and output its real-time position information. Single-object tracking is one of the important tasks in the field of computer vision, and its core is to continuously and accurately predict the position and size of the target in subsequent search frames. Current single-object tracking algorithms are prone to the problem of false target tracking when faced with situations such as target occlusion, rapid motion blur, or the target moving out of the field of view.
[0003] The problem of false target retention refers to the tracking result being distorted because the tracker, lacking effective target information, misidentifies background or similar targets as the tracked target. Existing solutions mainly determine whether the target is lost by setting a fixed confidence threshold, but this has significant limitations: First, the optimal threshold varies greatly in different scenarios (such as indoor / outdoor environments, changes in lighting, etc.), resulting in poor adaptability and limited applicability of the method; second, the confidence output of lightweight tracking algorithms lacks robustness, and even if the target has disappeared, the confidence level may still remain high, failing to meet the requirements of scenarios with low tolerance for false retention; third, this method relies on only a single confidence index, ignoring the target's motion patterns and category characteristics, making it prone to target drift or misjudgment in complex scenarios. Summary of the Invention
[0004] The purpose of this invention is to solve the above-mentioned technical problems existing in the prior art, and to provide a method and judgment system for eliminating false target delays in single target tracking. Through multi-dimensional feature complementary verification, the accuracy of target loss judgment and scene adaptability are improved, and efficient and stable single target tracking false delay elimination effect is achieved.
[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0006] A method for eliminating false target delays in single-target tracking, characterized by the following steps:
[0007] S1. Input target information:
[0008] Input the initial target template information and search frame;
[0009] The single-target tracking algorithm outputs the bounding box of the tracked target and the confidence score of the current frame based on the input initial template information of the target and the current frame.
[0010] S2. Determine the target status:
[0011] 1) The edge logic judgment module determines the location of the tracking target based on the bounding box of the tracking target in S1, and then sets different confidence thresholds according to different locations of the tracking target;
[0012] The confidence level of the current frame in S1 is compared with the confidence threshold, and the tracking target is determined based on the comparison result.
[0013] 2) The target classification and judgment module identifies the category of the target within the bounding box of the tracked target, and then compares the identified target category with the initial category of the tracked target. Based on the comparison result, it determines whether the tracked target has been lost.
[0014] 3) The motion constraint judgment module predicts the motion trajectory of the tracked target based on the motion state of the tracked target, compares the overlap between the motion trajectory and the bounding box of the tracked target in S1, and judges whether the tracked target is lost based on the comparison result.
[0015] S3. Comprehensive Judgment: Based on the judgment results of the edge logic judgment module, target classification judgment module, and motion constraint judgment module in S2, if any module determines that the tracked target is lost, then the tracked target is determined to be lost and the tracking box is eliminated.
[0016] Furthermore, in step S2, the edge logic judgment module divides the location of the tracked target into a central region and an edge region, and sets different confidence thresholds accordingly. The confidence threshold for the edge region is higher than that for the central region.
[0017] Furthermore, in step S2, the edge logic judgment module performs smoothing filtering on the current frame confidence in S1, and then compares the smoothed current frame confidence with the corresponding confidence threshold. If the current frame confidence is not greater than the confidence threshold, it is determined that the tracking target is lost.
[0018] Furthermore, the confidence score of the current frame is smoothed and filtered to obtain the average confidence score, which is the average confidence score of the current frame and the preset number of frames before it.
[0019] Furthermore, in step S2, the target classification judgment module uses a classification model to identify the category of the target within the bounding box of the tracked target. The size of the input image for the classification model is 64×64, and the step size of the last stage in the classification model is set to 1.
[0020] Furthermore, the classification model adopted is the RepVGG-A0 model.
[0021] Furthermore, in step S2, the target classification and judgment module uses a periodic verification method to identify the category of the target within the bounding box of the tracked target, and then performs smoothing filtering on the classification result.
[0022] Furthermore, during the initialization of the tracking template, the classification results of the template frame and the subsequent preset number of frames are calculated, and the category with the highest frequency is taken as the initial category of the tracking target; then, with a preset number of frames as a period, the classification results of the subsequent preset number of frames in each period are calculated, and the category with the highest frequency is taken as the current period category; if the current period category is different from the initial category, it is determined that the tracking target is lost.
[0023] Furthermore, in step S2, the motion constraint judgment module uses Kalman filtering to predict the bounding box of the tracked target in the current frame based on the historical position and size information of the tracked target, and calculates the intersection-union ratio (IUR) between the predicted bounding box and the bounding box of the tracked target in S1. If the IUR is less than the preset IUR threshold, the tracked target is determined to be lost.
[0024] A single-target tracking system for judging false stoppage of a target, characterized in that it includes: an edge logic judgment module, a target classification judgment module, and a motion constraint judgment module;
[0025] The edge logic judgment module is used to determine whether the tracking target is lost based on the bounding box and confidence score of the tracking target output by the search frame.
[0026] The target classification and judgment module is used to determine whether the tracked target has been lost based on changes in the target's category.
[0027] The motion constraint judgment module is used to determine whether the tracked target has been lost based on the degree of overlap between the location of the tracked target and its motion trajectory.
[0028] The present invention, by adopting the above-described technical solution, has the following beneficial effects:
[0029] 1. The edge logic judgment module makes judgments based on the location and confidence level of the tracked target, which is the simplest and most direct judgment method in single target tracking algorithms.
[0030] 2. The target classification and judgment module classifies the tracked target as a person in the initial frame. In the target loss frame, the target classification and judgment module judges the candidate targets output by the algorithm as background, thereby eliminating falsely detained targets.
[0031] 3. The motion constraint judgment module compares the position output by the single target tracking algorithm with the position predicted by the Kalman filter. If the deviation between the two exceeds the preset judgment threshold, it is determined that the tracked target is lost.
[0032] 4. This invention improves the accuracy and scene adaptability of target loss judgment through multi-dimensional feature complementarity verification, and achieves efficient and stable single target tracking error elimination effect. Attached Figure Description
[0033] The present invention will be further described below with reference to the accompanying drawings:
[0034] Figure 1 This is a flowchart of a single-target tracking target false delay elimination method according to the present invention;
[0035] Figure 2 This is a flowchart of the calculation process of the edge logic judgment module in this invention;
[0036] Figure 3 This is a flowchart of the calculation process for the target classification and judgment module in this invention;
[0037] Figure 4 This is a flowchart of the motion constraint judgment module in this invention.
[0038] Figure 5 This is a frame test result of the single-target tracking algorithm in this invention on the VisDrone validation set. Detailed Implementation
[0039] A single-target tracking system for determining whether a target is lost includes: an edge logic judgment module, a target classification judgment module, and a motion constraint judgment module. The edge logic judgment module is used to determine whether the tracked target is lost based on the bounding box and confidence level of the tracked target output in the search frame. The target classification judgment module is used to determine whether the tracked target is lost based on changes in the category of the tracked target. The motion constraint judgment module is used to determine whether the tracked target is lost based on the degree of overlap between the location of the tracked target and its motion trajectory.
[0040] like Figure 1 As shown, a method for eliminating false delays in single-target tracking includes the following steps:
[0041] S1. Input target information:
[0042] Input the initial target template information and the search frame, and denote the search frame as... ,in Let H represent the set of real numbers of dimension H×W, where H represents the height of the search frame image and W represents the width of the search frame image.
[0043] The single-target tracking algorithm outputs the bounding box of the tracked target and the confidence score of the current frame based on the input initial target template information and the current frame. The bounding box of the tracked target is denoted as... The confidence score of the current frame is recorded as Where l represents the horizontal coordinate of the top-left vertex of the bounding box, t represents the vertical coordinate of the top-left vertex of the bounding box, w represents the width of the bounding box, h represents the height of the bounding box, and T represents the transpose.
[0044] S2. Determine the target status:
[0045] 1) such as Figure 2As shown, the edge logic judgment module determines the location of the tracking target based on the bounding box of the tracking target in S1, and then divides the location of the tracking target into a central region and an edge region, and sets different confidence thresholds accordingly, so as to improve the ability to eliminate false target delays.
[0046] Generally, a target is more likely to disappear in the edge region than in the center region for the following reasons: First, compared to the center region, the target may leave the field of view in the edge region, making it impossible to capture the tracked target in the search frame; second, in the edge region, the search range of the tracked target will exceed the search frame, at which point the single target tracking algorithm will perform padding, resulting in a lower number of effective pixels in the search range than in the center region, thus affecting the accuracy of the tracking results.
[0047] Determining whether a target is in the edge region can be represented as:
[0048] .
[0049] in, , , This represents the coordinates of the bottom right corner of the target's bounding box. This represents the edge threshold, which is a constant.
[0050] The confidence threshold can be expressed as:
[0051] .
[0052] in, The confidence threshold representing the edge region. This represents the confidence threshold for the central region, and .
[0053] Next, the confidence score of the current frame in S1 is smoothed and filtered to obtain the average confidence score. The average confidence score is the average confidence score of the current frame and at most 5 frames before it, denoted as . .
[0054] average confidence With confidence threshold For comparison, if the average confidence level Not greater than the confidence threshold If so, the tracking target is determined to be lost.
[0055] 2) such as Figure 3 As shown, the target classification and judgment module uses a classification model to identify the category of the target within the bounding box of the tracked target. The classification model adopted is the RepVGG-A0 model, which is a lightweight open-source model that can reduce the impact on the frame rate of the single target tracking algorithm.
[0056] For small target tracking tasks, where the target occupies fewer pixels in the search frame, and to reduce classification processing latency, the size of the input image for the classification model is adjusted to 64×64. In addition, to avoid information loss due to excessively small deep feature map size, the step size of the last stage in the classification model is set to 1.
[0057] The classification model can be represented as:
[0058] .
[0059] in, This represents the probability that the classification model predicts the category to which the tracked target belongs.
[0060] The index with the highest calculated probability is taken as the category of the tracked target, represented as:
[0061] .
[0062] The total number of tracked target and background categories is C. When training the classification network, the training dataset contains target categories and background images, and the model parameters are optimized using the cross-entropy loss function.
[0063] The target classification module uses a periodic verification method to identify the category of targets within the bounding box of the tracked target, reducing the impact on the frame rate of the single-target tracking algorithm. Then, the classification results are smoothed and filtered to improve the stability of the target classification module. The process is as follows:
[0064] During the initialization of the tracking template, the classification results of the template frame and the next 4 frames are calculated, and the category with the highest frequency is taken as the initial category of the tracking target. Then, every 50 frames is considered as a cycle, and the classification results of the last 5 frames in each cycle are calculated, and the category with the highest frequency is taken as the current cycle category.
[0065] If the current period category is different from the initial category, the tracking target is determined to be lost.
[0066] Generally, when a single-object tracking algorithm mistakenly identifies the background or other types of objects as the tracking target, the object category within the bounding box will change.
[0067] 3) such as Figure 4 As shown, the motion constraint judgment module uses Kalman filtering to predict the motion trajectory of the tracked target based on the target's motion state.
[0068] Kalman filtering is a linear optimal recursive estimation algorithm that estimates the state information of the current time step based on the state of the previous time step.
[0069] Define the true state vector of the target position in the k-th frame as:
[0070] .
[0071] in, Represents the horizontal and vertical coordinates of the target. This indicates the target's speed in the horizontal and vertical directions.
[0072] To reduce computational complexity and impact on frame rate, this invention simplifies the Kalman filter by setting the control matrix to zero. The state transition equation can then be expressed as:
[0073] ,
[0074] .
[0075] in, The time interval between adjacent frames. The process noise of the k-th frame is represented by the following expression: Gaussian distribution, q is the process noise coefficient, I is the identity matrix, and F represents the state transition matrix.
[0076] The observation equation for tracking the target's position can be expressed as:
[0077] ;
[0078] .
[0079] in, Represents observation noise, obeying Gaussian distribution, , where r is the observation noise coefficient and H represents the observation transition matrix.
[0080] When initializing the target template, the velocity is set to 0, and the initial covariance matrix is set. In subsequent search frames, the position and state vector of the current frame and the covariance matrix are estimated using the position and state vector of the previous frame given by the single-target tracking algorithm, as follows:
[0081] , which represents the intermediate value of the current frame position state estimate;
[0082] , which represents the intermediate value of the current frame covariance matrix estimate.
[0083] Then read the observations of the current frame. And calculate the Kalman gain:
[0084] ,
[0085] .
[0086] Finally, the position state and covariance matrix are corrected and expressed as follows:
[0087] ,
[0088] .
[0089] Will As the target position state vector The Kalman filter estimate.
[0090] The process of predicting the size of the tracked target using Kalman filtering is similar to the process of predicting its position. We define the size state vector of the tracked target in the k-th frame as follows: .
[0091] in, Indicates the width and height of the target. This indicates the rate of change of width and height.
[0092] The state transition equation can be expressed as:
[0093] ,
[0094] .
[0095] in, The time interval between adjacent frames. The process noise of the k-th frame is represented by the following expression: Gaussian distribution, q is the process noise coefficient, I is the identity matrix, and F represents the state transition matrix.
[0096] The observation equation for tracking the target size can be expressed as:
[0097] ,
[0098] .
[0099] in, Represents observation noise, obeying Gaussian distribution, , where r is the observation noise coefficient and H represents the observation transition matrix.
[0100] When initializing the target template, the velocity is set to 0, and the initial covariance matrix is set. In subsequent search frames, the size state vector and covariance matrix of the current frame are estimated using the size state vector from the previous frame provided by the single-target tracking algorithm, as shown below:
[0101] ,
[0102] .
[0103] Then read the observations of the current frame. And calculate the Kalman gain:
[0104] ,
[0105] .
[0106] Finally, the size status and covariance matrix are corrected, and expressed as follows:
[0107] ,
[0108] .
[0109] Will As the target size state vector The Kalman filter estimate.
[0110] The predicted bounding box is then represented by the position and size estimates of the tracked target:
[0111] .
[0112] The bounding box obtained by the single-target tracking algorithm in S1 in the kth frame is represented as follows: Calculate the intersection-union ratio (IUR) of the bounding box obtained by the Kalman filter estimation with that obtained by the Kalman filter estimation. , is represented as:
[0113] .
[0114] If cross-comparison Less than the preset crossover ratio threshold If the result calculated by the single-target tracking algorithm differs significantly from the target's trajectory, then the tracked target is determined to be lost.
[0115] S3. Comprehensive Judgment: Based on the judgment results of the edge logic judgment module, target classification judgment module, and motion constraint judgment module in S2, if any module determines that the tracked target is lost, then the tracked target is determined to be lost and the tracking box is eliminated.
[0116] Figure 5The image shows a single-frame test result of the single-target tracking algorithm on the VisDrone validation set. In the figure, the box corresponding to gt (ground truth) represents the true location of the tracked target, and the other box represents the predicted location of the target output by the single-target tracking algorithm. The values marked above the boxes represent the prediction confidence. The test shows that when the tracked target is lost, the confidence score output by the algorithm is still as high as 0.607, significantly higher than the conventional discrimination threshold. Simple threshold judgment logic cannot identify target loss events in a timely manner, thus leading to the problem of false target retention. This problem is particularly prominent in scenarios with limited computing power, such as edge artificial intelligence. In such scenarios, lightweight algorithms or models are often used, and their accuracy limitations further increase the probability of false target retention.
[0117] The target classification and motion constraint judgment modules designed in this invention effectively solve the above-mentioned problems. The target classification module classifies the tracked target as a person in the initial frame. In the target loss frame, the target classification module classifies the candidate targets output by the algorithm as background, thereby eliminating falsely identified targets. The motion constraint judgment module compares the position output by the single-target tracking algorithm with the position predicted by the Kalman filter. If the deviation exceeds a preset discrimination threshold, the tracked target is determined to be lost. The edge logic judgment module makes a judgment based on the location and confidence level of the tracked target, which is the simplest and most direct judgment method in single-target tracking algorithms.
[0118] The above are merely specific embodiments of the present invention, but the technical features of the present invention are not limited thereto. Any simple changes, equivalent substitutions, or modifications made based on the present invention to solve essentially the same technical problems and achieve essentially the same technical effects are all covered within the protection scope of the present invention.
Claims
1. A method for eliminating false delays in single-target tracking, characterized in that... The process includes the following steps: S1, Input target information: Input the initial target template information and search frame; The single target tracking algorithm outputs the bounding box of the tracked target and the confidence score of the current frame based on the input initial target template information and the current frame; S2, Determine the target status: 1) The edge logic judgment module determines the location of the tracked target based on the bounding box of the tracked target in S1, and then sets different confidence thresholds according to different locations of the tracked target; The confidence score of the current frame in S1 is compared with the confidence threshold, and the tracking target is determined based on the comparison result; 2) The target classification and judgment module identifies the category of the target within the bounding box of the tracked target, and then compares the identified target category with the initial category of the tracked target. Based on the comparison result, it determines whether the tracked target has been lost. 3) The motion constraint judgment module predicts the motion trajectory of the tracked target based on the motion state of the tracked target, compares the overlap between the motion trajectory and the bounding box of the tracked target in S1, and judges whether the tracked target is lost based on the comparison result. S3. Comprehensive Judgment: Based on the judgment results of the edge logic judgment module, target classification judgment module, and motion constraint judgment module in S2, if any module determines that the tracked target is lost, then the tracked target is determined to be lost and the tracking box is eliminated.
2. The method for eliminating false delays in single-target tracking according to claim 1, characterized in that: In step S2, the edge logic judgment module divides the location of the tracked target into a central region and an edge region, and sets different confidence thresholds accordingly. The confidence threshold for the edge region is higher than that for the central region.
3. The method for eliminating false delays in single-target tracking according to claim 1, characterized in that: In step S2, the edge logic judgment module performs smoothing filtering on the current frame confidence in S1, and then compares the smoothed current frame confidence with the corresponding confidence threshold. If the current frame confidence is not greater than the confidence threshold, it is determined that the tracking target is lost.
4. The method for eliminating false delays in single-target tracking according to claim 3, characterized in that: The confidence score of the current frame is smoothed and filtered to obtain the average confidence score, which is the average confidence score of the current frame and the preset number of frames before it.
5. The method for eliminating false delays in single-target tracking according to claim 1, characterized in that: In step S2, the target classification and judgment module uses a classification model to identify the category of the target within the bounding box of the tracked target. The size of the input image for the classification model is 64×64, and the step size of the last stage in the classification model is set to 1.
6. The method for eliminating false delays in single-target tracking according to claim 5, characterized in that: The classification model used is the RepVGG-A0 model.
7. The method for eliminating false delays in single-target tracking according to claim 1, characterized in that: In step S2, the target classification and judgment module uses a periodic verification method to identify the category of the target within the bounding box of the tracked target, and then performs smoothing filtering on the classification result.
8. The method for eliminating false delays in single-target tracking according to claim 7, characterized in that: During the initialization of the tracking template, the classification results of the template frame and a preset number of subsequent frames are calculated, and the category with the highest frequency is used as the initial category of the tracking target. Then, taking a preset number of frames as a period, the classification results of the next preset number of frames in each period are calculated, and the category with the highest frequency is taken as the category of the current period. If the current period category is different from the initial category, the tracking target is determined to be lost.
9. The method for eliminating false delays in single-target tracking according to claim 1, characterized in that: In step S2, the motion constraint judgment module uses Kalman filtering to predict the bounding box of the tracked target in the current frame based on the historical position and size information of the tracked target. It calculates the intersection-union ratio (IUR) between the predicted bounding box and the bounding box of the tracked target in S1. If the IUR is less than the preset IUR threshold, the tracked target is determined to be lost.
10. A single-target tracking system for determining false delays, characterized in that, include: Edge logic judgment module, target classification judgment module, and motion constraint judgment module; The edge logic judgment module is used to determine whether the tracking target is lost based on the tracking target bounding box and confidence level output by the search frame. The target classification and judgment module is used to determine whether the tracked target has been lost based on changes in the category of the tracked target; The motion constraint judgment module is used to determine whether the tracked target has been lost based on the degree of overlap between the location of the tracked target and its motion trajectory.