A dynamic scene real-time denoising method for 3D point cloud labeling task

Abstract

The application discloses a kind of dynamic scene real-time denoising methods for 3D point cloud labeling task, comprising the following steps: S1, obtain continuous multiple frames of original 3D point cloud data under dynamic scene;S2, with the actual physical coverage range predicted as benchmark;S3, construct different neighborhood search structure respectively for three levels of area;S4, execute three levels of different denoising processing;S5, the current frame label frame is modified reversely consistent;S6, the label frame after modification is updated to label history track.The application is modified reversely to label frame by denoising result, and label frame is positively guided to subsequent denoising, form mutually promoted benign cycle, effectively maintain the contour integrity of dynamic target in continuous frame processing, simultaneously, through the area change trend monitoring of continuous frame label frame, realize the early warning and active compensation of label frame drift, through dynamic expansion edge protection range and weak feature point backfill suppress drift trend.

Description

Technical Field

[0001] This invention relates to the field of 3D point cloud data processing technology, and in particular to a real-time denoising method for dynamic scenes for 3D point cloud annotation tasks. Background Technology

[0002] In applications such as autonomous driving, robot environmental perception, 3D reconstruction, and high-precision map construction, 3D point cloud annotation is a core foundation for model training and algorithm verification. However, point cloud data continuously acquired by LiDAR or depth cameras in dynamic scenarios generally suffers from problems such as large data volume, significant inter-frame motion differences, uneven local density, dense noise at long distances, and prominent motion artifacts. Noisy points can directly cause annotation box offsets, missing targets, misclassification, and loss of key points, significantly reducing annotation accuracy and operational efficiency.

[0003] Existing 3D point cloud denoising methods are mostly based on static filtering, global statistical algorithms, or fixed thresholds, and generally adopt a one-way processing flow of denoising before annotation or referencing the current frame's annotation bounding box during denoising. For a long time, the industry has only focused on the noise removal rate and local feature preservation rate of single-frame denoising, but has completely ignored a hidden and fatal systemic flaw: there is a dynamic feedback deviation between the denoising result and the annotation bounding box, which can form an irreversible vicious cycle.

[0004] When the denoising algorithm incorrectly removes weak feature points on the edge of the labeled target, the annotator can only shrink the label box based on the incomplete point cloud. In the next frame of denoising, the algorithm will construct a smaller target area based on the shrunken label box, causing more edge points that originally belonged to the target to be included in the background area and removed.

[0005] This cycle repeats itself, causing small dynamic targets to gradually disappear in consecutive frames, and the bounding boxes to continuously drift and shrink, eventually leading to a systematic bias in the entire dataset. This problem is particularly prominent in high-speed dynamic scenes, low-quality point clouds, and small target annotations, which can cause serious missed detections and size estimation errors in the trained model. Existing denoising methods have not recognized the existence of this feedback mechanism, nor have they proposed an effective solution. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a real-time dynamic scene denoising method for 3D point cloud annotation tasks.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A real-time denoising method for dynamic scenes in 3D point cloud annotation tasks includes the following steps:

[0009] S1. Acquire raw 3D point cloud data of multiple consecutive frames in a dynamic scene, extract the current frame annotation box information and the historical trajectory data of the annotation boxes of the previous N frames, and predict the actual physical coverage of the target in the current frame based on motion consistency; at the same time, perform edge point trajectory inheritance, and incorporate the points belonging to the target edge in the previous frame into the target candidate region of the current frame through motion mapping.

[0010] S2. Based on the predicted actual physical coverage area, the point cloud is divided into three levels of regions: target core area, target edge buffer area, and background area. The target edge buffer area is a ring-shaped area that extends outward from the predicted range, and its width is dynamically adjusted according to the target's movement speed.

[0011] S3. Construct differentiated neighborhood search structures for the three-level regions respectively, calculate the three types of feature values ​​of density, motion and distance for each point and introduce regional weighting coefficients for correction, calculate feature distance based on the corrected feature values, and generate three-level denoising thresholds for the three regions.

[0012] S4. Perform three-level differentiated noise removal processing, using different noise removal intensities for different regions; perform secondary trajectory verification on all points identified as noise. If the point belongs to the target region and moves consistently in multiple consecutive frames in the historical trajectory, it is judged as a misjudgment and retained.

[0013] S5. Based on the denoised complete target point cloud, perform reverse consistency correction on the current frame bounding box, adjust its position and size to match the actual contour of the target; at the same time, detect the continuous shrinkage trend of the bounding box, and if there is a risk of drift, automatically trigger the edge compensation mechanism.

[0014] S6. Update the corrected bounding boxes to the historical trajectory and dynamically adjust the trajectory confidence based on the correction accuracy. Use this information for target coverage prediction and region segmentation in the next frame, forming a loop processing mechanism that optimizes both denoising and annotation.

[0015] Preferably, step S1 specifically involves: extracting the bounding box data of the first 5 frames, fitting the target motion trajectory based on Kalman filtering and predicting the actual coverage area of ​​the current frame; extracting edge points within a preset range of the bounding box edges in the first 3 frames, predicting their positions in the current frame based on the motion trajectory, and if there are points in the current frame that are less than 0.1m away from the predicted position, they are marked as historical inherited edge points and included in the target candidate region.

[0016] Preferably, the formula for calculating the width of the target edge buffer in step S2 is:

[0017]

[0018] in, For the buffer width, The target speed is defined as the speed of motion, and the maximum width of the buffer zone does not exceed 0.3m.

[0019] Preferably, the neighborhood search structure and weighting coefficients of the three-level regions in step S3 are configured as follows:

[0020] Target core area: k-nearest neighbor search, k=8-12, neighborhood radius is 0.4-0.6 times the global radius, weighting coefficient 1.2-1.8;

[0021] Target edge buffer: Hybrid neighborhood search, with a neighborhood radius of 0.8-1.2 times the global radius and a weighting coefficient of 0.8-1.2;

[0022] Background area: Voxelized nearest neighbor search, voxel size 0.2-0.3m, neighborhood radius 1.3-1.7 times the global radius, weighting coefficient 0.4-0.8.

[0023] Preferably, the formula for calculating the feature distance in step S3 is:

[0024]

[0025] in, The feature values ​​of the point to be judged are after correction. The eigenvalues ​​are the corrected values ​​for the neighborhood center points. The feature weights are denoted as , and the motion feature weights increase as the target velocity increases.

[0026] Preferably, the three-level denoising thresholds and elimination rules in step S4 are as follows: the target core area threshold is 1.6 times the base threshold, and only strong noise points are eliminated; the target edge buffer area threshold is 1.2 times the base threshold, and only noise points that deviate significantly from the trajectory are eliminated; the background area threshold is 0.8 times the base threshold, and strong denoising is performed.

[0027] Preferably, the rules for annotation box correction and drift compensation in step S5 are as follows: calculate the IOU between the smallest bounding rectangle of the denoised target and the original annotation box; if the IOU is less than 0.8, then correct the annotation box; if the cumulative decrease in the area of ​​the annotation box in 3 consecutive frames exceeds 15%, then expand the buffer width and replenish weak feature points.

[0028] Preferably, the confidence level is 1.0 when the IOU between the corrected bounding box and the target bounding rectangle is greater than 0.9 and 0.5 when it is less than 0.7; the prediction weight of the next frame is proportional to the confidence level of the historical trajectory.

[0029] The present invention has the following beneficial effects:

[0030] 1. This invention creates a virtuous cycle by using denoising results to correct bounding boxes in reverse and using the bounding boxes to guide subsequent denoising, effectively maintaining the contour integrity of dynamic targets in continuous frame processing. Simultaneously, by monitoring the trend of bounding box area changes across consecutive frames, it achieves early warning and proactive compensation for bounding box drift. By dynamically expanding the edge protection range and replenishing weak feature points, it suppresses drift trends, completing the shift from passive post-processing correction to proactive pre-processing prevention. This fundamentally solves the problem of target degradation and gradual disappearance during long-term annotation, significantly reducing the cost of manual annotation correction.

[0031] 2. This invention transmits and protects the verified target edge features from previous frames to the current frame through motion trajectory mapping, effectively compensating for edge feature loss caused by occlusion, distance attenuation, or noise interference in single-frame point clouds. Even if the target edge features are not obvious in the current frame, they can be preserved through historical information, significantly improving the labelability of incomplete targets. This allows targets that could not be labeled in a single frame to be directly used in the labeling process, expanding the applicability of the method in complex dynamic scenes.

[0032] 3. This invention employs a speed-adaptive dynamic edge buffer, making the buffer width linearly positively correlated with the target's motion speed. The protection range can be dynamically adjusted according to the target's motion state, adapting to the motion blur characteristics of point clouds at different motion speeds. Simultaneously, a secondary trajectory consistency verification step is added, introducing historical trajectory consistency constraints during noise discrimination. Points whose current frame features deviate but whose historical trajectories conform to the target's motion patterns are identified as weak feature points and retained. This design effectively removes background noise while fully preserving the target's edge features, significantly improving the target feature retention rate in low signal-to-noise ratio scenarios such as nighttime, rain, and fog, without requiring separate parameter configuration for different motion scenarios. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of a dynamic scene real-time denoising method for 3D point cloud annotation tasks proposed in this invention.

[0034] Figure 2 This is a flowchart illustrating the construction process of the three-layer neighborhood search structure proposed in this invention.

[0035] Figure 3 This is a flowchart of the three-level denoising threshold generation and differential denoising process proposed in this invention;

[0036] Figure 4 This is a flowchart of the annotation box correction and trajectory confidence update proposed in this invention. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0038] Reference Figure 1 A real-time denoising method for dynamic scenes in 3D point cloud annotation tasks includes the following steps:

[0039] S1. Acquire raw 3D point cloud data from multiple consecutive frames in a dynamic scene, extract the current frame bounding box information and the historical trajectory data of the bounding boxes from the previous N frames, and predict the actual physical coverage of the target in the current frame based on motion consistency; simultaneously, perform edge point trajectory inheritance, and incorporate points belonging to the target edge in the previous frame into the target candidate region of the current frame through motion mapping; specifically: extract the bounding box data from the previous 5 frames, fit the target motion trajectory based on Kalman filtering and predict the actual coverage of the current frame; extract edge points within a preset range of the bounding box edges in the previous 3 frames, predict their positions in the current frame based on the motion trajectory, and if there are points in the current frame whose distance from the predicted position is less than 0.1m, mark them as historically inherited edge points and incorporate them into the target candidate region.

[0040] S2, such as Figure 2 As shown, based on the predicted actual physical coverage area, the point cloud is divided into three hierarchical regions: the target core region, the target edge buffer zone, and the background region. The target edge buffer zone is a ring-shaped region extending outwards from the predicted range, and its width is dynamically adjusted according to the target's movement speed. The formula for calculating the width of the target edge buffer zone is:

[0041]

[0042] in, For the buffer width, The target speed is defined as the speed of motion, and the maximum width of the buffer zone does not exceed 0.3m.

[0043] S3. Construct differentiated neighborhood search structures for the three hierarchical regions respectively. Calculate three types of feature values—density, motion, and distance—for each point and introduce regional weighting coefficients for correction. Calculate the feature distance based on the corrected feature values ​​to generate the corresponding three-level denoising thresholds for the three regions. The neighborhood search structures and weighting coefficients for the three hierarchical regions are configured as follows:

[0044] Target core area: k-nearest neighbor search, k=8-12, neighborhood radius is 0.4-0.6 times the global radius, weighting coefficient 1.2-1.8;

[0045] Target edge buffer: Hybrid neighborhood search, with a neighborhood radius of 0.8-1.2 times the global radius and a weighting coefficient of 0.8-1.2;

[0046] Background area: Voxelized nearest neighbor search, voxel size 0.2-0.3m, neighborhood radius 1.3-1.7 times the global radius, weighting coefficient 0.4-0.8.

[0047] The formula for calculating the feature distance is:

[0048]

[0049] in, The feature values ​​of the point to be judged are after correction. The eigenvalues ​​are the corrected values ​​for the neighborhood center points. The feature weights are denoted as , and the motion feature weights increase as the target velocity increases.

[0050] S4, such as Figure 3 As shown, a three-level differentiated denoising process is performed, with different noise removal intensities applied to different regions. For all points identified as noise, a secondary trajectory verification is performed. If a point belongs to the target region and exhibits consistent motion across multiple consecutive frames in the historical trajectory, it is considered a misjudgment and retained. The three-level denoising thresholds and removal rules are as follows: the target core region threshold is 1.6 times the base threshold, removing only strong noise points; the target edge buffer zone threshold is 1.2 times the base threshold, removing only noise points that significantly deviate from the trajectory; and the background region threshold is 0.8 times the base threshold, performing strong denoising.

[0051] S5, such as Figure 4 As shown, based on the denoised complete target point cloud, the current frame's bounding box undergoes reverse consistency correction, adjusting its position and size to match the actual target contour. Simultaneously, the continuous shrinkage trend of the bounding box is detected, and if there is a risk of drift, the edge compensation mechanism is automatically triggered. The bounding box correction and drift compensation rules are as follows: calculate the IOU between the minimum bounding rectangle of the denoised target and the original bounding box; if the IOU is less than 0.8, the bounding box is corrected; if the area of ​​the bounding box in three consecutive frames decreases by more than 15% cumulatively, the buffer width is expanded and weak feature points are replenished.

[0052] S6. Update the corrected bounding boxes to the historical trajectory and dynamically adjust the trajectory confidence based on the correction accuracy. This confidence is used for target coverage prediction and region segmentation in the next frame, forming a loop processing mechanism that optimizes both denoising and annotation. The trajectory confidence update rule is as follows: the confidence is 1.0 when the IOU between the corrected bounding box and the target's bounding rectangle is greater than 0.9, and 0.5 when it is less than 0.7. The prediction weight for the next frame is proportional to the historical trajectory confidence.

[0053] Example 1: High-speed dynamic vehicle-mounted LiDAR point cloud annotation scenario

[0054] This embodiment is applied to an autonomous driving scenario on an urban expressway. A 16-line LiDAR is used to collect point clouds at a frame rate of 10Hz and a vehicle speed of 55–65 km / h. The labeled objects are dynamic small targets such as pedestrians, non-motorized vehicles, and small motor vehicles. Existing conventional denoising methods and the method of this invention are used to process the same continuous 100 frames of point cloud data. Key statistical indicators are shown in the table below.

[0055] Table 1: Comparison of the effects of conventional noise reduction methods and the method of this invention in high-speed scenarios.

[0056] Evaluation indicators Conventional noise reduction methods Method of the present invention Target edge point retention rate 76.2% 89.5% consecutive frame bounding box drift 0.48m 0.29m Target progressive contraction incidence 82% 63% Frequency of manual secondary correction 18 times per 100 frames 7 times per 100 frames

[0057] As can be clearly seen from Table 1, compared with existing conventional methods, this invention relies on motion trajectory prediction, dynamic edge buffer and trajectory secondary verification mechanism to greatly improve the ability to retain edge points of high-speed moving targets; effectively suppress continuous frame annotation box drift, basically eliminate the phenomenon of dynamic targets gradually shrinking and becoming smaller as the number of frames increases, significantly reduce the amount of manual post-correction work, and adapt to the continuous annotation operation requirements under high-speed dynamic road conditions.

[0058] Example 2: Low-speed complex occlusion annotation scenario for indoor robots

[0059] This embodiment is applied to indoor office and building corridor scenarios, using a depth camera to collect 3D point clouds. The scene is characterized by pedestrian movement and occlusion, sparse local point clouds, and high levels of noise from nearby points. The main labeled targets are pedestrians and movable facilities. Similarly, 100 frames were continuously sampled, and the existing methods were compared with the method of this invention. The key indicators are shown in the table below:

[0060] Table 2: Comparison of the effects of conventional noise reduction methods and the method of this invention in indoor low-speed occlusion scenarios.

[0061] Evaluation indicators Conventional noise reduction methods Method of the present invention Success rate can be marked on the target after occlusion. 65.8% 81.4% False rejection rate of weak feature edge points 28.5% 15.3% Does parameter tuning and adaptability need to be adjusted? Parameters need to be adjusted for each scenario. No manual parameter adjustment Completeness of damaged target repair 61.3% 79.6%

[0062] As can be clearly seen from Table 2, existing methods rely solely on single-frame point cloud features for judgment, which easily leads to the accidental deletion of weak feature points in occluded areas, making it difficult to properly label incomplete targets. This invention, through edge point trajectory inheritance and three-layer regional differential protection, can utilize historical frame information to complete the contour features missing in the current frame due to occlusion, significantly reducing the accidental removal of weak feature points. It eliminates the need for manual parameter configuration for indoor scenes, significantly improving the completeness and usability of target labeling under occlusion conditions.

[0063] Example 3: High-precision annotation scene for large-scale outdoor 3D reconstruction

[0064] This embodiment is applied to 3D reconstruction and annotation scenarios of outdoor parks and building facades. It uses a gimbal-mounted LiDAR for long-distance acquisition, covering a wide point cloud area with large target size range and high background clutter noise, requiring high precision in contour integrity and inter-frame annotation consistency. 100 frames of typical point cloud data were selected for comparative testing, and the key statistical indicators are shown in the table below.

[0065] Table 3: Comparison of the effects of conventional denoising methods and the method of this invention in external 3D reconstruction scenarios.

[0066] Evaluation indicators Conventional noise reduction methods Method of the present invention Building and target outline completeness 72.6% 84.1% Effective background noise removal rate 78.3% 92.8% Inter-frame bounding box position consistency 70.5% 86.2% 3D reconstruction model deformation probability 35.7% 14.2%

[0067] As clearly shown in Table 3, existing methods, which use a uniform threshold and fixed region division, struggle to balance background noise removal with the preservation of large structure edges. This invention, through three-level differentiated thresholds, reverse correction of annotation boxes, and dynamic updating of trajectory confidence, efficiently removes stray noise from the outdoor background while fully preserving the contour details of buildings and various targets. The consistency of inter-frame annotations is significantly improved, effectively reducing the local deformation problem of the 3D reconstruction model caused by annotation drift, and meeting the standards for high-precision 3D annotation and reconstruction operations.

[0068] In summary, this invention addresses the problems of existing 3D point cloud annotation denoising methods that only use single-frame unidirectional processing and ignore inter-frame correlation, which easily lead to annotation box drift and shrinkage, loss of target edge features, and poor consistency of continuous annotations. It proposes a real-time denoising method based on trajectory prediction, region layering, differentiated denoising, and annotation closed-loop feedback.

[0069] Furthermore, this invention relies on historical labeled trajectories to predict the actual coverage area of ​​the target, sets an edge buffer that adapts to the movement speed, and matches a differentiated neighborhood structure and hierarchical denoising threshold through three-layer region division. At the same time, it introduces a trajectory consistency secondary verification mechanism to distinguish between real noise and weak feature points of the target, and uses the complete point cloud after denoising to reverse-calibrate the labeled box, and combines the trajectory confidence dynamic update to form a closed-loop linkage between frames.

[0070] Compared to traditional fixed-parameter, single-threshold denoising methods, this invention eliminates the need for manual parameter tuning for different scenarios. It effectively preserves the dynamic target edge contours, suppresses continuous frame annotation drift and progressive target shrinkage, and is suitable for diverse application environments such as high-speed vehicle scenarios, low-speed indoor occlusion scenarios, and large-scale outdoor 3D reconstruction. While ensuring effective noise filtering, it significantly improves the integrity of point cloud annotations and inter-frame stability, reduces manual correction costs, and has good engineering practicality and scenario adaptability.

[0071] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A real-time denoising method for dynamic scenes in 3D point cloud annotation tasks, characterized in that, Includes the following steps: S1. Acquire raw 3D point cloud data of multiple consecutive frames in a dynamic scene, extract the current frame annotation box information and the historical trajectory data of the annotation boxes of the previous N frames, and predict the actual physical coverage of the target in the current frame based on motion consistency; at the same time, perform edge point trajectory inheritance, and incorporate the points belonging to the target edge in the previous frame into the target candidate region of the current frame through motion mapping. S2. Based on the predicted actual physical coverage area, the point cloud is divided into three levels of regions: target core area, target edge buffer area, and background area. The target edge buffer area is a ring-shaped area that extends outward from the predicted range, and its width is dynamically adjusted according to the target's movement speed. S3. Construct differentiated neighborhood search structures for the three-level regions respectively, calculate the three types of feature values ​​of density, motion and distance for each point and introduce regional weighting coefficients for correction, calculate feature distance based on the corrected feature values, and generate three-level denoising thresholds for the three regions. S4. Perform three-level differentiated noise reduction processing, using different noise removal intensities for different areas; Perform secondary trajectory verification on all points identified as noise. If the point belongs to the target area and moves consistently in multiple consecutive frames in the historical trajectory, it is considered a misjudgment and is retained. S5. Based on the denoised complete target point cloud, perform reverse consistency correction on the current frame bounding box, adjust its position and size to match the actual contour of the target; at the same time, detect the continuous shrinkage trend of the bounding box, and if there is a risk of drift, automatically trigger the edge compensation mechanism. S6. Update the corrected bounding boxes to the historical trajectory and dynamically adjust the trajectory confidence based on the correction accuracy. Use this information for target coverage prediction and region segmentation in the next frame, forming a loop processing mechanism that optimizes both denoising and annotation.

2. The method for real-time denoising of dynamic scenes for 3D point cloud annotation tasks according to claim 1, characterized in that, Step S1 specifically involves: extracting the bounding box data from the first 5 frames, fitting the target motion trajectory based on Kalman filtering, and predicting the actual coverage area of ​​the current frame; extracting edge points within a preset range of the bounding box edges in the first 3 frames, predicting their positions in the current frame based on the motion trajectory, and marking any points in the current frame that are less than 0.1m away from the predicted position as historical inherited edge points and including them in the target candidate region.

3. The method for real-time denoising of dynamic scenes for 3D point cloud annotation tasks according to claim 1, characterized in that, The formula for calculating the width of the target edge buffer in step S2 is: in, For the buffer width, The target speed is defined as the speed of motion, and the maximum width of the buffer zone does not exceed 0.3m.

4. The method for real-time denoising of dynamic scenes for 3D point cloud annotation tasks according to claim 1, characterized in that, The neighborhood search structure and weighting coefficients for the three-level regions in step S3 are configured as follows: Target core area: k-nearest neighbor search, k=8-12, neighborhood radius is 0.4-0.6 times the global radius, weighting coefficient 1.2-1.8; Target edge buffer: Hybrid neighborhood search, with a neighborhood radius of 0.8-1.2 times the global radius and a weighting coefficient of 0.8-1.2; Background area: Voxelized nearest neighbor search, voxel size 0.2-0.3m, neighborhood radius 1.3-1.7 times the global radius, weighting coefficient 0.4-0.

8.

5. A real-time denoising method for dynamic scenes in 3D point cloud annotation tasks according to claim 1, characterized in that, The formula for calculating the feature distance in step S3 is: in, The feature values ​​of the point to be judged are after correction. The eigenvalues ​​are the corrected values ​​for the neighborhood center points. The feature weights are denoted as , and the motion feature weights increase as the target velocity increases.

6. A method for real-time denoising of dynamic scenes for 3D point cloud annotation tasks according to claim 1, characterized in that, In step S4, the three-level noise reduction threshold and elimination rules are as follows: the target core area threshold is 1.6 times the base threshold, and only strong noise points are eliminated; The target edge buffer threshold is 1.2 times the base threshold, removing only noise points that significantly deviate from the trajectory; the background area threshold is 0.8 times the base threshold, performing strong noise reduction.

7. A real-time denoising method for dynamic scenes in 3D point cloud annotation tasks according to claim 1, characterized in that, The rules for annotation box correction and drift compensation in step S5 are as follows: calculate the IOU between the smallest bounding rectangle of the denoised target and the original annotation box; if the IOU is less than 0.8, then correct the annotation box. If the area of ​​the bounding box decreases by more than 15% cumulatively over three consecutive frames, the buffer width is increased and weak feature points are replenished.

8. The method according to claim 1, characterized in that, The trajectory confidence update rule in step S6 is as follows: when the IOU between the corrected bounding box and the target bounding rectangle is greater than 0.9, the confidence is 1.0, and when it is less than 0.7, the confidence is 0.5; the prediction weight of the next frame is proportional to the historical trajectory confidence.

Images