Trajectory reconstruction method for vulnerable road users based on multi-dimensional features of aerial videos

Abstract

The present invention discloses a method for reconstructing a target trajectory based on multi-dimensional features of aerial video, which belongs to the field of digital image detection technology, and aims to solve the problem in the prior art that it is difficult to accurately restore the real motion trajectory of the target due to interference from multiple factors. It includes: S1, data labeling and target detection of vulnerable road users based on aerial video; S2, generating a trajectory map to be completed for vulnerable road users according to time and position information; S3, selecting relevant trajectories, and calculating the matching degree between the relevant trajectories and the trajectories to be completed to obtain a completed trajectory; S4, respectively performing logic checks and continuity checks on multiple completed trajectories to obtain multiple alternative trajectories; S5, obtaining the best trajectory through a machine learning model, and completing the trajectory map to be completed. The present invention significantly optimizes the detection of targets as small targets through the trajectory reconstruction method.

Description

Technical Field

[0001] The present invention relates to the technical field of digital image detection, and in particular to a method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video. Background Art

[0002] In the urban traffic system, signalized intersections play a key role as important nodes of cross flow. However, at these busy intersections, vulnerable road users (including pedestrians, cyclists, moped riders, motorcyclists, etc.) often face more complex safety challenges. Compared with motor vehicles, these road participants are usually more vulnerable and lack the protection of vehicles, making their traffic safety at signalized intersections a focus of attention.

[0003] As small targets, vulnerable road users face several unique challenges in detection. The limited size of the target, uneven sample distribution and high target density make the resolution low, which increases the difficulty of labeling detection. On the other hand, the existing public datasets are mainly for large / medium-sized targets, and there is a lack of large-scale datasets for small target detection. This makes the detection of vulnerable road users of great practical significance. Since vulnerable road users may perform complex nonlinear movements such as acceleration, deceleration, and turning, the trajectory reconstruction using simple linear models becomes inaccurate. In addition, the observed data may be affected by factors such as sensor noise and environmental interference, resulting in inaccuracy of the trajectory. The presence of noise makes it difficult to accurately restore the true motion trajectory of the target. In addition, it is also very important to choose the right motion model for different types of targets, such as pedestrians, cyclists, moped riders, and motorcyclists.

[0004] In view of the above problems, the present invention proposes a method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial videos. Summary of the invention

[0005] The purpose of the present invention is to provide a method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial videos, so as to solve the technical problem in the prior art that it is difficult to accurately restore the real movement trajectory of vulnerable road users due to interference from multiple factors.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video provided by the present invention comprises the following steps:

[0008] S1. Collect aerial videos of vulnerable road users, and perform data labeling and target detection on vulnerable road users based on the aerial videos;

[0009] S2. acquiring time and position information of trajectory data of vulnerable road users by detecting targets in the aerial video, and generating a trajectory map to be completed of the vulnerable road users according to the time and position information;

[0010] S3, selecting multiple related trajectories that are similar to the trajectory to be completed, and calculating the matching degree between the related trajectories and the trajectory to be completed, to obtain multiple completed trajectories;

[0011] S4, performing logic check and continuity check on the multiple completion trajectories in S3 to obtain multiple candidate trajectories;

[0012] S5. Select the best trajectory from multiple candidate trajectories through a machine learning model, and use the best trajectory to complete the trajectory map to be completed in S2.

[0013] Compared with other sensors, aerial video provides a global bird's-eye view, which can more comprehensively capture the movement trajectory of vulnerable road users at intersections. By deeply understanding the movement patterns of vulnerable road users, potential traffic conflict concentration points can be identified, so as to take preventive measures and reduce the possibility of accidents. The position of small targets is identified and located through the target detection algorithm in the aerial video image frame, and the deep learning model is used for target detection. Various algorithms are used to track the position of the target in continuous video image frames. Trajectory reconstruction combines the position information of the target at different time steps to form the movement trajectory of the target in space, which usually includes interpolation, smoothing or filtering operations on the target position data. Vulnerable road users appear as detection targets in each video image frame, and trajectory reconstruction solves the problem of missed detection and false detection due to interference from the target tracking algorithm and other factors.

[0014] Optionally or preferably, the method of performing data labeling and target detection on vulnerable road users in step S1 includes:

[0015] A dataset of vulnerable road users is obtained by annotating the acquired aerial videos, a vulnerable road user tracking model is constructed based on the dataset, and trajectory information of vulnerable road users is extracted according to the vulnerable road user tracking model;

[0016] The YOLO V8 detection model is used to obtain the bounding box information of vulnerable road users in the current aerial video image frame, and the position of vulnerable road users in the next frame of aerial video image is predicted based on the Kalman filter.

[0017] Optionally or preferably, the method for calculating the degree of matching between the relevant trajectory and the trajectory to be completed in step S3 includes matching calculation based on the shortest distance, matching calculation based on a motion model, and matching calculation based on target image features.

[0018] Optionally or preferably, the method based on the shortest distance matching calculation comprises the following steps:

[0019] A1. Calculate the Euclidean distance d between the relevant trajectory and the target frame in the same frame:

[0020]

[0021] Among them, (x1, y1) is the position coordinate of the target box in the relevant trajectory, and (x2, y2) is the position coordinate of the target box in the trajectory to be completed;

[0022] A2. The Euclidean distance of each frame between the relevant trajectory and the trajectory to be completed is used to form a loss matrix. Each element (i, j) of the loss matrix represents the Euclidean distance from the relevant trajectory i to the trajectory to be completed j. The best trajectory match is obtained through the Hungarian matching algorithm as the completed trajectory.

[0023] Optionally or preferably, the method based on motion model matching calculation includes the following steps:

[0024] B1. Calculate the velocity v, acceleration a and angular velocity ω of the target vulnerable road user based on the driving distance and time difference between the previous and next frames in the aerial video of the target vulnerable road user:

[0025]

[0026]

[0027]

[0028] Where (x, y, w, h, θ) is the position information of the bounding box of the current frame, (x', y', w', h', θ') is the position information of the bounding box of the next frame, and (x", y", w", h", θ") is the position information of the bounding box of the previous frame; (x, y), (x', y') and (x", y") are the center positions of the bounding box, (w, h), (w', h') and (w", h") are the width and height of the bounding box respectively, θ, θ' and θ" are the angular directions of the bounding box; Δt is the time interval between two adjacent frames;

[0029] B2. Perform kinematic feature matching on the target vulnerable road user, set thresholds for speed, acceleration and angular velocity, and determine whether the similarity between the relevant trajectory and the trajectory to be completed meets the requirements;

[0030] If the speed, acceleration or angular velocity calculated from the coordinates of the two frames before and after the aerial video of the target vulnerable road user is greater than a threshold of one, the match is abandoned;

[0031] If the speed, acceleration, and angular velocity calculated from the coordinates of the two frames before and after the aerial video of the target vulnerable road user are less than a threshold of one, they are associated into a trajectory as a completed trajectory.

[0032] Optionally or preferably, the method for performing matching calculation based on target image features comprises the following steps:

[0033] C1. Extract features from the bounding box of the target vulnerable road user and set the pixel area S of the vulnerable road user displayed in the aerial video to:

[0034] S=W×H

[0035] Where W is the width of the bounding box, and H is the height of the bounding box;

[0036] C2. Setting a second threshold value for pixel area, if the pixel area of ​​the same type of vulnerable road user is smaller than the second threshold value, then giving up the match;

[0037] If the pixel area of ​​vulnerable road users of the same type is greater than threshold 2, they are associated into a trajectory as a completed trajectory.

[0038] Optionally or preferably, the method of performing logic checking on the plurality of completed trajectories in step S4 is analyzing the trajectory pattern of the vulnerable road user in the completed trajectories;

[0039] The method for analyzing the trajectory pattern of the vulnerable road user in the completed trajectory is to draw a position jump distribution map, and check whether there is a position change of the vulnerable road user in the adjacent frames of the completed trajectory that does not conform to the actual situation according to the position jump distribution map;

[0040] If the position of the vulnerable road user in the adjacent frames of the completed trajectory does not conform to the actual situation, the matching is abandoned;

[0041] If the vulnerable road user does not have any position changes that are inconsistent with the actual situation in the adjacent frames of the completed trajectory, it will be used as an alternative trajectory.

[0042] Optionally or preferably, the method of performing continuity check on the plurality of completed trajectories in step S4 includes performing position continuity check and time continuity check on the vulnerable road user trajectories;

[0043] Among them, the method of checking the position continuity of the completed trajectory is to determine whether the position jump between adjacent frames of the completed trajectory is greater than a normal value;

[0044] If the position jump between adjacent frames of the completed trajectory is greater than the normal value, the match is abandoned;

[0045] If the position jump between adjacent frames of the completed trajectory is smaller than the normal value, it will be used as an alternative trajectory.

[0046] Optionally or preferably, the method of checking the time continuity of the completed trajectory is to determine whether the time interval between adjacent frames of the completed trajectory is stable;

[0047] If the time interval between adjacent frames of the completed trajectory is unstable, the matching is abandoned;

[0048] If the time interval between adjacent frames of the completed trajectory is stable, it will be used as an alternative trajectory.

[0049] Optionally or preferably, in step S5, the method of selecting the best completion trajectory from multiple candidate trajectories through the machine learning model is:

[0050] Through the machine learning model, the features of each candidate trajectory are used as input, and the labeled trajectory quality is used as output, and the candidate trajectory with the best trajectory quality is selected as the best trajectory.

[0051] Based on the above technical solution, the present invention can at least produce the following technical effects:

[0052] (1) The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial videos provided by the present invention realizes comprehensive modeling of the mobility behavior of vulnerable road users by integrating multi-dimensional information from aerial videos. It also successfully improves the accuracy and robustness of the trajectory of vulnerable road users by using computer vision and machine learning technologies.

[0053] (2) The detection of vulnerable road users as small targets is significantly optimized through trajectory reconstruction methods;

[0054] (3) The present invention helps to better cope with complex urban environments and diverse behaviors of vulnerable road users, and enhances the credibility of trajectory reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention; it is obvious that the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments, and all other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work belong to the scope of protection of the present invention.

[0057] Example

[0058] See also Figure 1 , a method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video, comprising the following steps:

[0059] S1. Collecting aerial videos of vulnerable road users, and performing data labeling and target detection on vulnerable road users based on the aerial videos;

[0060] In this embodiment, the method for data labeling and target detection of vulnerable road users is to obtain a data set of vulnerable road users by labeling the acquired aerial video, build a vulnerable road user tracking model based on the data set, and extract trajectory information of vulnerable road users according to the vulnerable road user tracking model;

[0061] The YOLO V8 detection model is used to obtain the bounding box information of vulnerable road users in the current aerial video image frame, and the position of vulnerable road users in the next frame of aerial video image is predicted based on the Kalman filter.

[0062] Target detection focuses on individual video image frames, providing independent target position information by identifying and locating the target object in each frame; the continuity of video image frames provides a temporal relationship for target tracking, so that targets in adjacent frames can be associated to form motion trajectories.

[0063] Target tracking not only analyzes the associated targets between the previous and next frames and assigns a unique identifier (ID) to each target, but also lays the foundation for subsequent trajectory analysis and trajectory prediction. This unique identifier remains consistent during the movement of the target, allowing the system to track and understand the temporal dynamics of the target, providing deeper information for video analysis. Therefore, it is necessary to perform data association on candidate targets in multiple frames before and after and obtain multiple motion trajectories.

[0064] S2. acquiring time and position information of trajectory data of vulnerable road users by detecting targets in the aerial video, and generating a trajectory map to be completed of the vulnerable road users according to the time and position information;

[0065] In this embodiment, vulnerable road users refer to pedestrians, cyclists, moped riders and motorcyclists. By using the AI ​​model to infer attribute information, including the location and time information of the trajectory data, the system generates a trajectory map to be completed, showing the trajectory of the target (vulnerable road user) over time. Because vulnerable road users are affected by various factors, they may lose their location information in a certain period of time, resulting in discontinuous trajectories, so subsequent completion processing is required.

[0066] S3, selecting multiple related trajectories that are similar to the trajectory to be completed, and calculating the matching degree between the related trajectories and the trajectory to be completed, to obtain multiple completed trajectories;

[0067] The system detects the presence of disconnected points in the blank trajectory in the coordinate diagram of the trajectory to be completed. The system handles this problem by calculating the similarity between the disconnected points of the blank trajectory and the approximate trajectory in the historical trajectory. Since the same target in adjacent frames in the video sequence has similar features, including the distance between two adjacent frames is within a certain range, the kinematic features of the target conform to certain rules, have similar image features, meet the same trajectory logic, and the motion trajectory is continuous.

[0068] Based on these features, we select several similar trajectories (i.e., related trajectories) to calculate the matching degree with the trajectory to be completed, and finally retain the completed trajectory with the highest matching degree (i.e., the best trajectory). This method helps to improve the continuity and accuracy of the trajectory, allowing the system to more effectively deal with the situation of blank trajectories.

[0069] The above-mentioned method for calculating the matching degree includes matching calculation based on the shortest distance, matching calculation based on the motion model and matching calculation based on the target image features;

[0070] The method based on the shortest distance matching calculation includes the following steps:

[0071] A1. Each track has a target box position. For the same frame target box, calculate the Euclidean distance d between the relevant track and the track to be completed:

[0072]

[0073] Among them, (x1, y1) is the position coordinate of the target box in the relevant trajectory, and (x2, y2) is the position coordinate of the target box in the trajectory to be completed;

[0074] A2. The Euclidean distance of each frame between the relevant trajectory and the trajectory to be completed is used to form a loss matrix. Each element (i, j) of the loss matrix represents the Euclidean distance from the relevant trajectory i to the trajectory to be completed j. The best trajectory match is found through the Hungarian matching algorithm and used as the completed trajectory to minimize the overall matching loss.

[0075] The method based on motion model matching calculation comprises the following steps:

[0076] B1. Calculate the velocity v, acceleration a and angular velocity ω of the target vulnerable road user based on the driving distance and time difference between the previous and next frames in the aerial video of the target vulnerable road user:

[0077]

[0078]

[0079]

[0080] Where (x, y, w, h, θ) is the position information of the bounding box of the current frame, (x', y', w', h', θ') is the position information of the bounding box of the next frame, and (x", y", w", h", θ") is the position information of the bounding box of the previous frame; (x, y), (x', y') and (x", y") are the center positions of the bounding box, (w, h), (w', h') and (w", h") are the width and height of the bounding box respectively, θ, θ' and θ" are the angular directions of the bounding box; Δt is the time interval between two adjacent frames;

[0081] B2. Perform kinematic feature matching on the target vulnerable road user, set thresholds for speed, acceleration and angular velocity, and determine whether the similarity between the relevant trajectory and the trajectory to be completed meets the requirements;

[0082] If the speed, acceleration or angular velocity calculated from the coordinates of the two frames before and after the aerial video of the target vulnerable road user is greater than the threshold value 1, the matching is abandoned;

[0083] If the speed, acceleration and angular velocity calculated from the coordinates of the two frames before and after in the aerial video of the target vulnerable road user are less than the threshold value one, they are associated into a trajectory as the completed trajectory.

[0084] The method for performing matching calculation based on target image features includes the following steps:

[0085] C1. Extract features from the bounding box of the target vulnerable road user and set the pixel area S of the vulnerable road user displayed in the aerial video to:

[0086] S=W×H

[0087] Where W is the width of the bounding box, and H is the height of the bounding box;

[0088] C2. Setting a second threshold value for pixel area, if the pixel area of ​​the same type of vulnerable road user is smaller than the second threshold value, then abandoning the match;

[0089] If the pixel areas of vulnerable road users of the same type are larger than the second threshold, they are associated into a trajectory as the completed trajectory.

[0090] S4, performing logic check and continuity check on the multiple completion trajectories described in S3 to obtain multiple candidate trajectories;

[0091] In this embodiment, the method for performing logic checking is to analyze the trajectory pattern of the vulnerable road user in the completed trajectory; the method for performing continuity checking includes performing position continuity checking and time continuity checking on the trajectory of the vulnerable road user; this process helps to ensure the accuracy of tracking results, improve the reliability of target recognition and classification, and reduce the probability of false alarms and misdetections, which is also a requirement that must be met for trajectory reconstruction.

[0092] Visualization and statistical analysis tools are used to evaluate the logic of trajectories and the continuity of trajectories, and trajectory paths are plotted to observe the movement continuity of vulnerable road users in time and space.

[0093] Analyzing trajectory patterns is used to ensure that the target conforms to its expected behavior. By drawing a distribution map of position jumps, it is checked whether there is a large position change of vulnerable road users in adjacent frames of the completed trajectory, and the movement of the target trajectory is consistent with the actual situation of the scene; for example, pedestrian trajectories should move in pedestrian areas, and vehicle trajectories should follow road rules. For example, if the target moves to the left in one frame and suddenly moves to the right in the next frame, further inspection of the data or algorithm may be required.

[0094] If the position of the vulnerable road user in the adjacent frames of the completed trajectory does not conform to the actual situation, the matching is abandoned;

[0095] If the vulnerable road user does not have any position changes that are inconsistent with the actual situation in the adjacent frames of the completed trajectory, it will be used as an alternative trajectory.

[0096] The method for checking the position continuity of the trajectory is to determine whether the position jump between adjacent frames of the completed trajectory is greater than a normal value; if there is a large position change between adjacent frames, further analysis may be required to determine whether this is a reasonable movement.

[0097] If the position jump between adjacent frames of the completed trajectory is greater than the normal value, the match is abandoned;

[0098] If the position jump between adjacent frames of the completed trajectory is smaller than the normal value, it will be used as an alternative trajectory.

[0099] The method of checking the temporal continuity of the trajectory is to determine whether the time interval between adjacent frames of the completed trajectory is stable; abnormal time intervals may indicate problems with data acquisition or system failures.

[0100] If the time interval between adjacent frames of the completed trajectory is unstable, the matching is abandoned;

[0101] If the time interval between adjacent frames of the completed trajectory is stable, it will be used as an alternative trajectory.

[0102] S5. Select the best trajectory from multiple candidate trajectories through a machine learning model, and use the best trajectory to complete the trajectory map to be completed in S2.

[0103] In this embodiment, after calculating the shortest distance, kinematic features (motion model) and image features of the relevant trajectories, the trajectories that do not meet the conditions are screened out after logical checks and continuity tests. Through the machine learning model, the features of each trajectory are used as input, and the quality of the annotated trajectory is used as output. Then, the new trajectory is predicted by the model to determine which trajectory is more credible, and finally the trajectory that best meets the reality is selected as the best trajectory, and the best trajectory is selected to complete the blank trajectory.

[0104] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video, characterized in that: The following steps are involved: S1. Collecting aerial videos of vulnerable road users, and performing data labeling and target detection on vulnerable road users based on the aerial videos; S2. acquiring time position information of the trajectory data of the vulnerable road user by detecting the target in the aerial video, and generating a trajectory map to be completed of the vulnerable road user according to the time position information; S3, selecting a plurality of related trajectories similar to the trajectory to be completed, and calculating the matching degree between the related trajectories and the trajectory to be completed, to obtain a plurality of completed trajectories; S4, performing logic check and continuity check on the multiple completion trajectories described in S3 to obtain multiple candidate trajectories; The method of performing logic check on the plurality of the completed trajectories in step S4 is to analyze the trajectory pattern of the vulnerable road user in the completed trajectories; The method for analyzing the trajectory pattern of the vulnerable road user in the completed trajectory is to draw a position jump distribution map, and check whether there is a position change of the vulnerable road user in the adjacent frames of the completed trajectory that does not conform to the actual situation according to the position jump distribution map; If the position of the vulnerable road user in the adjacent frames of the completed trajectory does not conform to the actual situation, the matching is abandoned; If the vulnerable road user does not have any position changes that are inconsistent with the actual situation in the adjacent frames of the completed trajectory, it will be used as an alternative trajectory; S5. Select the best trajectory from multiple candidate trajectories through a machine learning model, and use the best trajectory to complete the trajectory map to be completed in S2.

2. The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video according to claim 1, characterized in that: The method for data labeling and target detection of vulnerable road users in step S1 includes: A dataset of vulnerable road users is obtained by annotating the acquired aerial video, a vulnerable road user tracking model is constructed based on the dataset, and trajectory information of vulnerable road users is extracted according to the vulnerable road user tracking model; The YOLO V8 detection model is used to obtain the bounding box information of vulnerable road users in the current aerial video image frame, and the position of vulnerable road users in the next frame of the aerial video image is predicted based on the Kalman filter.

3. The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video according to claim 1, characterized in that: The method for calculating the matching degree between the relevant trajectory and the trajectory to be completed in step S3 includes matching calculation based on the shortest distance, matching calculation based on the motion model, and matching calculation based on the target image features.

4. The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video according to claim 3, characterized in that: The method based on the shortest distance matching calculation comprises the following steps: A1. Calculate the Euclidean distance between the relevant trajectory and the target frame in the same frame between the trajectory to be completed d : in,( x 1, y 1) is the position coordinate of the target box in the relevant trajectory, ( x 2, y 2) The position coordinates of the target box in the trajectory to be completed; A2. The Euclidean distance of each frame between the relevant trajectory and the trajectory to be completed forms a loss matrix. Each element of the loss matrix ( i , j ) represents the relevant trajectory i To the track to be completed j The Euclidean distance of is used to obtain the best trajectory match through the Hungarian matching algorithm, and the best trajectory is used as the completed trajectory.

5. The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video according to claim 3, characterized in that: The method based on motion model matching calculation comprises the following steps: B1. Calculate the speed of the target vulnerable road user based on the driving distance and time difference of the previous and next frames in the aerial video of the target vulnerable road user v , acceleration a and angular velocity ω : in, is the position information of the current frame bounding box, is the position information of the next frame bounding box, is the position information of the bounding box of the previous frame; , and is the center position of the bounding box, , and are the width and height of the bounding box, respectively. , and is the angular direction of the bounding box; is the time interval between two adjacent frames; B2. Perform kinematic feature matching on the target vulnerable road user, set thresholds for speed, acceleration and angular velocity, and determine whether the similarity between the relevant trajectory and the trajectory to be completed meets the requirements; If the speed, acceleration or angular velocity calculated from the coordinates of the two frames before and after the aerial video of the target vulnerable road user is greater than the threshold value 1, the matching is abandoned; If the speed, acceleration and angular velocity calculated from the coordinates of the two frames before and after in the aerial video of the target vulnerable road user are less than the threshold value one, they are associated into a trajectory as the completed trajectory.

6. The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video according to claim 3, characterized in that: The method for performing matching calculation based on target image features comprises the following steps: C1. Extract features from the bounding box of the target vulnerable road user and set the pixel area of ​​the vulnerable road user displayed in the video from the aerial perspective S for: in, W is the width of the bounding box, H is the height of the bounding box; C2. Setting a second threshold value for pixel area, if the pixel area of ​​the same type of vulnerable road user is smaller than the second threshold value, then abandoning the match; If the pixel areas of vulnerable road users of the same type are larger than the second threshold, they are associated into a trajectory as the completed trajectory.

7. The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video according to claim 1, characterized in that: The method of performing continuity check on the plurality of the completed trajectories in step S4 comprises performing position continuity check and time continuity check on the vulnerable road user trajectories; Among them, the method of checking the position continuity of the completed trajectory is to determine whether the position jump between adjacent frames of the completed trajectory is greater than a normal value; If the position jump between adjacent frames of the completed trajectory is greater than the normal value, the match is abandoned; If the position jump between adjacent frames of the completed trajectory is smaller than the normal value, it will be used as an alternative trajectory.

8. The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video according to claim 7, characterized in that: The method of checking the temporal continuity of the completed trajectory is to determine whether the time interval between adjacent frames of the completed trajectory is stable; If the time interval between adjacent frames of the completed trajectory is unstable, the matching is abandoned; If the time interval between adjacent frames of the completed trajectory is stable, it will be used as an alternative trajectory.

9. The method for reconstructing the trajectory of vulnerable road users based on multi-dimensional features of aerial video according to claim 1, characterized in that: In step S5, the method of selecting the best completion trajectory from multiple candidate trajectories through the machine learning model is: Through the machine learning model, the features of each candidate trajectory are used as input, and the labeled trajectory quality is used as output, and the candidate trajectory with the best trajectory quality is selected as the best trajectory.

Images