Track stabilization method for queuing vehicles at intersection based on 4D millimeter wave radar

Abstract

The application relates to a method for realizing track stabilization of queuing vehicles at an intersection based on a 4D millimeter wave radar, wherein the method comprises the following steps: acquiring 4D millimeter wave radar point cloud data and road calibration information; pre-processing the millimeter wave radar point cloud data according to the intersection calibration information; performing DBSCAN clustering on the pre-processed 4D millimeter wave point cloud; associating the generated track with the clustered bounding box, and generating a new track on the unassociated clustered box; after the track enters the queuing area, realizing queuing area track management according to the track index of two frames of front and rear lanes, and then performing K-Means clustering on the point cloud in the queuing area according to the track index of the current frame of lanes, associating the track in the queuing area with the clustering after the clustering, and outputting stable track information.

Description

Technical Field

[0001] The present invention belongs to the technical field of 4D millimeter-wave radar, and specifically provides a method for stabilizing the tracks of vehicles queuing at intersections based on 4D millimeter-wave radar. Background Art

[0002] In the field of intelligent transportation, the development of traffic perception technology is crucial. It not only provides real-time data for traffic management but also supports intelligent and safe driving of vehicles. With advances in artificial intelligence, the Internet of Things, and sensor technologies, traffic perception technology has made significant progress, especially in multi-dimensional perception capabilities, such as vehicle, road, and air. Among them, 4D millimeter-wave radar, as a key traffic perception tool, plays an increasingly important role in intelligent transportation systems due to its wide monitoring range, long-range detection capabilities, beyond-line-of-sight perception, and strong anti-interference capabilities.

[0003] At traffic intersections, 4D millimeter-wave radar multi-target tracking can provide effective vehicle location and speed information, and can also be used to calculate the length of vehicles queued at the intersection and traffic flow. Track stability is one of the keys to successful target tracking, and maintaining track stability for queued vehicles at intersections is a key issue for 4D millimeter-wave radar. Traditional millimeter-wave radar multi-target tracking can output stable tracks for moving targets. However, when target vehicles enter the queue area at the intersection, the number of vehicle targets increases and they approach each other, resulting in a decrease in speed or even zero speed. The corresponding point clouds of the vehicles cannot be distinguished, resulting in a decrease in multi-target tracking performance. Tracking accuracy and stability are difficult to guarantee, and track jumps and false targets occur frequently, which cannot meet the requirements of multi-target tracking. Therefore, the core difficulty in maintaining track stability for queued vehicles lies in how to cluster the millimeter-wave radar point clouds corresponding to the tracks into clusters corresponding to the tracks in the absence of speed dimension information. Summary of the Invention

[0004] The present invention proposes a method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter-wave radar.

[0005] The technical solution for achieving the purpose of the present invention is: a method for stabilizing the track of vehicles queuing at an intersection based on a 4D millimeter-wave radar, comprising:

[0006] Step 1: Obtain 4D millimeter-wave radar point cloud data and road calibration information;

[0007] Step 2: Preprocess the millimeter-wave radar point cloud data based on road calibration information;

[0008] Step 3: Perform DBSCAN clustering on the preprocessed 4D millimeter wave point cloud to obtain a clustering box containing the center coordinate position, length, width, height, and average radial velocity;

[0009] Step 4: Associate the generated tracks with the clustering boxes, and generate new tracks for the unassociated clustering boxes;

[0010] Step 5: Determine whether the generated track enters the queue area. If it does, perform queue area track management, i.e. jump to step 6. If it does not enter the queue area, jump to step 7.

[0011] Step 6: After the track enters the queueing area, the queueing area track is managed based on the lane track indexes of the previous and next two frames. Then, K-Means clustering is performed on the queueing area point cloud based on the lane track index of the current frame. After clustering, the queueing area track is associated with the cluster.

[0012] Step 7: Output stable track information.

[0013] Preferably, the specific method of preprocessing the millimeter wave radar point cloud data according to the road calibration information in step 2 is:

[0014] The region of interest (ROI) is obtained from road calibration. The region of interest includes the specific location coordinates of motor vehicle lanes, non-motor vehicle lanes, and intersections. The specific location coordinates of intersections use a left-handed coordinate system. The 4D millimeter-wave radar is used as the coordinate origin. The point cloud of the 4D millimeter-wave radar is filtered, and point cloud preprocessing is achieved by setting thresholds for the height and speed attributes of the non-ROI area.

[0015] Preferably, in step 3, DBSCAN clustering is performed on the preprocessed 4D millimeter wave point cloud, and a specific method for outputting a bounding box including the center coordinate position, length, width, height, and average radial velocity is as follows:

[0016] DBSCAN clustering is performed based on the three-dimensional coordinate position and relative radial velocity in the preprocessed millimeter-wave radar point cloud information. A radar point cloud is selected as the core point, and the difference in three-dimensional coordinate position and relative radial velocity between this point and the adjacent radar point cloud is calculated to see if it meets the clustering parameters. If the clustering parameters are met, the adjacent radar points are judged as qualified adjacent points, and the number of adjacent points is calculated. After finding the adjacent points of the radar point cloud, the adjacent points are used as the core points to continue looking for adjacent points. The steps of finding adjacent points are repeated until all adjacent points are found, and a cluster containing n points is obtained.

[0017] According to the characteristics of millimeter-wave radar point cloud, two sets of clustering parameters are set with 80m of road as the boundary. The specific parameters within 80m are:

[0018]

[0019] 80m away:

[0020]

[0021] Among them, x core ,y core , z core Indicates the three-dimensional distance of the core point relative to the millimeter wave radar, vr core Indicates the relative radial velocity of the core point relative to the millimeter wave radar, x border ,y border , z border Indicates the three-dimensional distance of the nearby point relative to the millimeter wave radar, vr core Indicates the relative radial velocity of the nearby point relative to the millimeter-wave radar;

[0022] Select the maximum and minimum values ​​in the x direction in the cluster, that is, x max ,x min , and the maximum and minimum values ​​in the y direction, that is, y max ,y min , use these four points to get the cluster bounding box;

[0023] The center coordinate x of the bounding box cneter =(x max +x min ) / 2,y cneter =(y max +y min ) / 2, the length of the bounding box is length = x max -x min , width = y max -y min , find the maximum and minimum values ​​in the z direction in the cluster frame, that is, z max ,z min , height=z max -z min , add up the relative radial velocities of all radar points in the cluster frame to get vr total , the average radial velocity vr is obtained according to the number of radar points in the cluster box mean =vr total / n.

[0024] Preferably, step 4 associates the generated track with the clustered point cloud clusters. The specific method for generating new tracks from the unassociated clusters is as follows:

[0025] Perform global matching on clusters and tracks, calculate the association scores between clusters and tracks, select a group of clusters with the largest association scores and tracks for association; for clusters without corresponding tracks, create a new track.

[0026] Preferably, the association score between clusters and tracks is composed of a weighted combination of distance score, area score, and speed score, and the calculation process is:

[0027] The distance score is calculated according to the following formula:

[0028]

[0029] Among them, x c is the distance of the clustering box in the x direction, x t is the distance of the track in the x direction, y c is the distance of the clustering box in the y direction, y t is the distance of the track in the y direction, dist cri is the preset distance threshold;

[0030] The area score is calculated according to the following formula:

[0031]

[0032] Among them, area c is the area of ​​the cluster box, area t is the area of ​​the track, min represents the minimum value, and max represents the maximum value;

[0033] The speed score is calculated according to the following formula:

[0034]

[0035] Among them, VR c is the average radial velocity of the cluster box, vr t is the average radial velocity of the track, vr cri is the preset speed threshold;

[0036] The association score is calculated according to the following formula:

[0037] score total =0.4×score distance +0.2×score area +0.4×score velocity .

[0038] Preferably, in step 6, after the track enters the queuing area, the track management of the queuing area is implemented according to the lane track indexes of the previous and next two frames, and then the queuing area point cloud is clustered according to the lane track index of the current frame. After clustering, the specific method of associating the queuing area track with the cluster is as follows:

[0039] Step 6.1: When the track enters the queuing area, a lane track index is created and the lane track index number of the current frame is stored to achieve mapping from track to lane;

[0040] Step 6.2: Search the lane track index of the previous frame and compare it with the lane track index of the current frame to perform lost track query;

[0041] Step 6.3: Check the lost track target. If the track is valid, recall the lost target track and repeat step 6.1.

[0042] Step 6.4: Search the lane track index of the current frame and compare it with the lane track index of the previous frame to find the newly added track;

[0043] Step 6.5: Verify the newly added track. If the track is invalid, delete the false track and repeat step 6.1.

[0044] Step 6.6: Perform K-Means clustering on the queuing area point cloud;

[0045] Step 6.7: Associate the queuing area tracks with the clusters.

[0046] Compared with the existing technology, the present invention has the following significant advantages: the present invention solves the problem of unstable tracking track of queuing vehicles at intersections in traffic scenarios, forms a stable track before entering the queuing area, and recalls or deletes the track based on the comparison of the previous and next frames after entering the queuing area, thereby realizing track management of the queuing area at the intersection; through the managed track, K-Means clustering is performed on the radar point cloud of the queuing area, and then the queuing area is associated, so that the queuing area track can be associated with the corresponding point cloud, reconstructing the observation of the queuing area track, and realizing the stability of the queuing area track.

[0047] The present invention will be described in further detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flow chart of a method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter wave radar according to the present invention.

[0049] Figure 2 It is a schematic diagram of the queuing area track management process.

[0050] Figure 3 This is a schematic diagram of the queue area clustering process. DETAILED DESCRIPTION

[0051] The following describes the specific implementation of the embodiment of the present invention in detail with reference to the accompanying drawings. It should be understood that the specific implementation described herein is only used to illustrate and explain the embodiment of the present invention and is not used to limit the embodiment of the present invention.

[0052] See also Figure 1As shown, a method for stabilizing the track of vehicles queuing at intersections based on 4D millimeter-wave radar is designed to use the point cloud provided by the 4D millimeter-wave radar to generate a stable prior track by continuously tracking moving targets that have not entered the queue area. When a moving target enters the queue area, the lane track management of the queue area is performed, and the track of the vehicle in the queue area is accurately matched with the clustered point cloud, achieving accurate tracking and track stabilization of the queued vehicles. The method comprises the following steps:

[0053] Step 1: Obtain 4D millimeter-wave radar point cloud data and road calibration information;

[0054] Step 2: Preprocess the millimeter-wave radar point cloud data based on road calibration information;

[0055] Step 3: Perform DBSCAN clustering on the preprocessed 4D millimeter wave point cloud and output the cluster box containing the center coordinate position, length, width, height and average radial velocity;

[0056] Step 4: Associate the generated tracks with the clustering boxes, and generate new tracks for the unassociated clustering boxes;

[0057] Step 5: Determine whether the generated track enters the queue area. If it does, perform queue area track management, i.e. jump to step 6. If it does not enter the queue area, jump to step 7.

[0058] Step 6: After the track enters the queueing area, the queueing area track is managed based on the lane track indexes of the previous and next two frames. Then, K-Means clustering is performed on the queueing area point cloud based on the lane track index of the current frame. After clustering, the queueing area track is associated with the cluster.

[0059] Step 7: Output stable track information.

[0060] As a preferred embodiment of the present invention, the road calibration information includes the yaw angle, pitch angle, roll angle, radar height, and specific location coordinates of the motor vehicle lane, non-motor vehicle lane and intersection.

[0061] The 4D millimeter-wave radar is set at the intersection.

[0062] As a preferred embodiment of the present invention, step 2 is specifically as follows:

[0063] The region of interest (ROI), i.e., the specific location coordinates of motor vehicle lanes, non-motor vehicle lanes, and intersections, is obtained from road calibration. The location coordinates use a left-handed coordinate system with the 4D millimeter-wave radar as the coordinate origin. Based on this information, the point cloud of the 4D millimeter-wave radar is filtered. Point cloud preprocessing is achieved by setting thresholds for non-ROI areas, height, speed, and other attributes. When the coordinates of the radar point cloud are in the non-ROI area or the point cloud height in the ROI area is greater than 5m and the point cloud speed is between 0 and 1m / s, an invalid mark is given to the point, and these point clouds are not used in subsequent steps to achieve filtering.

[0064] As a preferred embodiment of the present invention, the step 3 is specifically as follows:

[0065] DBSCAN clustering is performed based on the three-dimensional coordinate position and relative radial velocity in the preprocessed millimeter-wave radar point cloud information. A radar point cloud is selected as the core point, and the difference in three-dimensional coordinate position and relative radial velocity between this point and the adjacent radar point cloud is calculated to see if it meets the clustering parameters. If the clustering parameters are met, the adjacent radar points are judged as qualified adjacent points, and the number of adjacent points is calculated. After finding the adjacent points of the radar point cloud, the adjacent points are used as core points to continue looking for adjacent points. This step is repeated until all adjacent points are found, resulting in a cluster containing n points.

[0066] Based on the characteristics of millimeter-wave radar point clouds, two sets of clustering parameters are set with 80m of road as the boundary. When the distance in the x direction of the radar point cloud is less than 80m, the following clustering parameters are used:

[0067]

[0068] When the distance in the x-direction of the radar point cloud is greater than 80m, the following clustering parameters are used:

[0069]

[0070] Among them, x core ,y core , z core Indicates the three-dimensional distance of the core point relative to the millimeter wave radar, vr core Indicates the relative radial velocity of the core point relative to the millimeter wave radar, x border ,y border , z border Indicates the three-dimensional distance of the nearby point relative to the millimeter wave radar, vr core Indicates the relative radial velocity of the nearby point relative to the millimeter-wave radar.

[0071] Select the maximum and minimum values ​​in the x direction in the cluster, that is, x max ,x min, and the maximum and minimum values ​​in the y direction, that is, y max ,y min , use these four points to get the cluster bounding box. The center coordinates of the bounding box are x cneter =(x max +x min ) / 2,y cneter =(y max +y min ) / 2, the length of the bounding box = x max -x min , width=y max -y min , find the maximum and minimum values ​​in the z direction in the cluster frame, that is, z max ,z min , height=z max -z min , add up the relative radial velocities of all radar points in the cluster frame to get vr total , according to the number of radar points in the cluster box, the average radial velocity vr is obtained mean =vr total / n.

[0072] As a preferred embodiment of the present invention, step 4 is specifically as follows:

[0073] Perform global matching on clustering frames and tracks, and calculate the association score between clustering frames and tracks. The association score is composed of the weighted distance score, area score and speed score. Select the set of clustering frames and tracks with the largest association score to associate; for clustering frames without corresponding tracks, create a new track.

[0074] The distance score is calculated according to the following formula:

[0075]

[0076] Among them, x c is the distance of the clustering box in the x direction, x t is the distance of the track in the x direction, y c is the distance of the clustering box in the y direction, y t is the distance of the track in the y direction, dist cri is the preset distance threshold.

[0077] The area score is calculated according to the following formula:

[0078]

[0079] Among them, area c is the area of ​​the cluster box, area t is the area of ​​the track, min represents the minimum value, and max represents the maximum value.

[0080] The speed score is calculated according to the following formula:

[0081]

[0082] Among them, VR c is the average radial velocity of the cluster box, vr t is the average radial velocity of the track, vr cri is the preset speed threshold.

[0083] The association score is calculated according to the following formula:

[0084] score total =0.4×score distance +0.2×score area +0.4×score velocity ;

[0085] As a preferred embodiment of the present invention, step 5 is specifically as follows:

[0086] Determine whether the generated track enters the queue area. If the center point coordinate of the current frame track is less than the distance of the preset queue area, that is, x t <x q , where x t is the distance of the track in the x direction, x q To preset the distance of the queuing area, perform queuing area track management and jump to step 6. If you have not entered the queuing area, jump to step 7.

[0087] As a preferred embodiment of the present invention, step 6 is specifically as follows:

[0088] Step 6.1: When the track enters the queuing area, a lane track index is established and the lane track index number of the current frame is stored to achieve mapping from track to lane.

[0089] As a preferred embodiment of the present invention, step 6.1 is specifically as follows:

[0090] When the target track enters the queuing area, that is, the coordinate x of the target track meets the preset range gate, then x track –x cross <d cri , where x track is the distance of the track in the x direction of the kth frame, x cross The distance of the intersection in the x direction, d cri The target has formed a stable track. A new lane track index is created and the target track index number and the corresponding lane number are stored in it. Each lane is stored in the order of the track coordinate x distance to the intersection.

[0091] Step 6.2: Search the lane track index of the previous frame and compare it with the lane track index of the current frame to perform lost track query;

[0092] As a preferred embodiment of the present invention, the step 6.2 is specifically as follows:

[0093] According to the lane track index of the kth frame, starting from the first target of the i-th lane, search for the lane track index of the k+1th frame. If the track target is not found, search for the track target in the lanes i–1 and i+1. If the target is not found, it is marked as lost track, where i is the lane number of the road calibration information;

[0094] Step 6.3: Check the lost track target. If the track is valid, recall the lost target track and repeat step 6.1.

[0095] As a preferred embodiment of the present invention, the step 6.3 is specifically as follows:

[0096] Check lost targets, set thresholds for the number of detection frames and speed of lost tracks, and determine targets below the threshold as false targets to reduce false tracks formed by ghost points in the millimeter-wave radar point cloud. Check the distance between the lost target and the stop line to determine whether the first vehicle in the team has left. If the first vehicle in the team has left, check whether there is target movement in the current frame. If there is target movement, do not modify it; otherwise, keep the target to prevent the first vehicle in the team from being lost. Recall lost tracks and add the track attribute values ​​back to track management.

[0097] Step 6.4: Search the lane track index of the current frame and compare it with the lane track index of the previous frame to find the newly added track.

[0098] As a preferred embodiment of the present invention, the step 6.4 is specifically as follows:

[0099] Based on the lane track index of the k+1th frame, starting from the first target in the i-th lane, search for the lane track index of the k-th frame in the track of the lane track index. If not found, it is determined that a new track appears in the i-th lane of the k+1th frame. Then check whether the i-1 and i+1 lanes of the k-th frame have the same track. If not, it is determined that the track is a new track, where i is the lane number of the road calibration information.

[0100] Step 6.5: Check the newly added track. If the track is invalid, delete the false track and repeat step 6.1. As a preferred embodiment of the present invention, the step 6.5 is specifically as follows:

[0101] Check the newly added tracks, set thresholds for the detection frame number and speed of the newly added tracks, and judge them as false tracks if they are less than the thresholds, thereby reducing the false tracks formed by ghost points in the millimeter-wave radar point cloud; delete the newly added tracks and remove them from track management.

[0102] Step 6.6: Perform K-Means clustering on the queuing area point cloud;

[0103] As a preferred embodiment of the present invention, Figure 3 As shown, the queue area clustering in step 6.6 is as follows:

[0104] The number of lane tracks is obtained from the lane track index of the current frame. The 4D millimeter-wave radar point cloud of the queuing area is clustered by lane. K-Means clustering is performed on the millimeter-wave radar point cloud of the current lane. The number of clusters is set to the number of lane tracks k. K points are selected from the current lane point cloud as cluster centers, and the Euclidean distances of all points to these k points are calculated, that is, Among them, x c is the distance of the cluster center in the x direction, x i is the distance of the i-th radar point in the x direction, y c is the distance of the cluster center in the y direction, y i is the distance of the i-th radar point in the y direction, and the point with the minimum Euclidean distance from the cluster center is assigned to the cluster center. After all points are assigned, k clusters are formed; the mean of each cluster is calculated and the k cluster centers are updated, that is, Among them, x c is the distance of the cluster center in the x direction, x j is the distance of the jth radar point in the cluster in the x direction, y c is the distance of the cluster center in the x direction, and y is the distance of the jth radar point in the y direction within the cluster. Then calculate the difference between the new cluster center and the previous cluster center. When the difference is less than 0.0001 or the number of iterations reaches 9, the iteration ends and the point cloud clustering of the current lane is completed. Repeat this step until the point clouds of all lanes are clustered.

[0105] Select the maximum and minimum values ​​in the x direction in the cluster, that is, x max ,x min , and the maximum and minimum values ​​in the y direction, that is, y max ,y min , use these four points to get the cluster bounding box. The center coordinates of the bounding box are x cneter =(x max +x min ) / 2,y cneter =(y max +y min) / 2, the length of the bounding box = x max -x min , width=y max -y min , find the maximum and minimum values ​​in the z direction in the cluster frame, that is, z max ,z min , height=z max -z min , add up the relative radial velocities of all radar points in the cluster frame to get vr total , according to the number of radar points in the cluster box, the average radial velocity vr is obtained mean =vr total / n.

[0106] Step 6.7: Associate the queuing area tracks with the clusters.

[0107] As a preferred embodiment of the present invention, the step 6.7 is specifically as follows:

[0108] Match the clustering frames and tracks of the first lane by lane, and calculate the association score of the lane clustering frames and tracks. The association score is composed of the sum of the distance score, area score and speed score. The specific calculation method is the same as step 4. Select the group of clustering frames and tracks with the largest association score and associate them until the clustering frames and tracks of all lanes are associated.

[0109] Among them, the lane track management of steps 6.1 to 6.5 is as follows: Figure 2 shown.

[0110] In actual application, the method for stabilizing the track of vehicles queuing at intersections based on 4D millimeter-wave radar has the following main processes:

[0111] (1) Input: 4D millimeter-wave radar point cloud, road calibration information;

[0112] (2) Frame k: The target that has not entered the queue area forms a stable track through multi-frame association. When the track enters the queue area, due to the decrease in target speed, the target corresponding point cloud cannot be clustered into the corresponding target point cloud cluster using DBSCAN. The point clouds of the front and rear targets in the same lane will be clustered into one cluster, or the point clouds of the targets in adjacent lanes will be clustered into one cluster, resulting in problems such as track jumps or false tracks in the queue area. Therefore, when the track enters the queue area, the lane track index is used to record the lane index number entering the queue area as the comparison group for the lane track management of the k+1 frame. When the target track enters the queue area, that is, the coordinate x of the target track meets the preset distance gate, then x track –x cross <d cri , where x trackis the distance of the track in the x direction of the kth frame, x cross The distance of the intersection in the x direction, d cri The target has formed a stable track. Create a new array and store the target track index and the corresponding lane number in it. Each lane is stored in the order of the track coordinate x distance to the intersection.

[0113] (3) Frame k+1: First, record the lane track index of the k+1 frame entering the queue area and compare it with the lane track index of the previous frame; according to the lane track index of the kth frame, start from the first target of the i-th lane and search in the track of the lane track index of the k+1th frame. If the track target is not found, search for the track target in the lanes i–1 and i+1. If the target is not found, it is marked as a lost track, where i is the lane number of the road calibration information; check the lost target, set a threshold for the number of detection frames and speed of the lost track, and judge it as a false target if it is less than the threshold, reduce the false tracks formed by the ghost points of the millimeter wave radar, check the distance of the lost target from the stop line, and determine whether it is the first car in the queue. If the first car in the queue leaves, check whether there is a target in the current frame. If the target moves, do not modify it; otherwise, keep the target to prevent the team leader from being lost. Lost tracks are recalled and their attribute values ​​are added back to track management. Based on the lane track index of the k+1th frame, compare the lane track index of the i-th lane in the k+1th frame with the lane track index of the kth frame. If a new track appears in lane i in the k+1th frame, check whether the track exists in lanes i-1 and i+1 of the kth frame. If not, the track is considered a new track, where i is the lane number in the road calibration information. New tracks are checked, and thresholds are set based on the number of detection frames and speed of the new track. Tracks below the threshold are considered false tracks, reducing false tracks caused by ghost points of the millimeter-wave radar. New tracks are deleted and removed from track management. Finally, the lane track index of the k+1th frame is updated.

[0114] (4) Frame k+1: Based on the lane track index of frame k+1, the number of tracks k for lane i is obtained. K-Means clustering is performed on the 4D millimeter-wave radar point cloud for that lane. The number of clusters is set to the number of tracks in the current lane, and the clustering parameter uses the Euclidean distance. This continues until all lanes are clustered. Next, the lane tracks are associated with the clusters in the queuing area.

[0115] The method for stabilizing the tracks of vehicles queuing at intersections based on 4D millimeter-wave radar of the present invention solves the problem of unstable tracks of queuing vehicles at intersections when using 4D millimeter-wave radar target tracking in traffic scenarios. A stable track is formed before entering the queuing area. After entering the queuing area, the track is recalled or deleted based on the comparison of the previous and next frames, thereby realizing track management of the queuing area at the intersection. The managed track is used to perform K-Means clustering on the radar point cloud of the queuing area, and then the queuing area is associated, so that the track of the queuing area can be associated with the corresponding point cloud, and the observation of the track of the queuing area is reconstructed to achieve the stability of the track of the queuing area.

[0116] The present invention significantly improves the accuracy of traffic monitoring, enhances the safety of vehicle driving, and optimizes traffic flow management, thereby contributing to the development of intelligent transportation systems.

[0117] The above-described embodiments merely represent preferred implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that those skilled in the art will be able to make numerous variations and improvements without departing from the spirit of the present invention, and these variations and improvements are fully within the scope of protection of the present invention.

Claims

1. A method for achieving track stabilization of vehicles queuing at intersections based on 4D millimeter-wave radar, characterized in that: include: Step 1: Obtain 4D millimeter-wave radar point cloud data and road calibration information; Step 2: Preprocess the millimeter-wave radar point cloud data based on road calibration information; Step 3: Perform DBSCAN clustering on the preprocessed 4D millimeter wave point cloud to obtain a clustering box containing the center coordinate position, length, width, height, and average radial velocity; Step 4: Associate the generated tracks with the clustering boxes, and generate new tracks for the unassociated clustering boxes; Step 5: Determine whether the generated track enters the queue area. If it does, perform queue area track management, i.e. jump to step 6. If it does not enter the queue area, jump to step 7. Step 6: After the track enters the queue area, the queue area track management is implemented based on the lane track indexes of the previous and next two frames. Then, based on the lane track index of the current frame, K-Means clustering is performed on the queue area point cloud. After clustering, the queue area track is associated with the cluster. The specific method is as follows: Step 6.1: When the track enters the queuing area, a lane track index is created and the lane track index number of the current frame is stored to achieve mapping from track to lane; Step 6.2: Search the lane track index of the previous frame and compare it with the lane track index of the current frame to perform lost track query; Step 6.3: Check the lost track target. If the track is valid, recall the lost target track and repeat step 6.

1. Step 6.4: Search the lane track index of the current frame and compare it with the lane track index of the previous frame to find the newly added track; Step 6.5: Verify the newly added track. If the track is invalid, delete the false track and repeat step 6.

1. Step 6.6: Perform K-Means clustering on the queuing area point cloud; Step 6.7: Associate the queuing area tracks with the clusters; Step 7: Output stable track information.

2. The method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter-wave radar according to claim 1, characterized in that: The specific method of step 2 for preprocessing millimeter wave radar point cloud data according to road calibration information is as follows: The region of interest (ROI) is obtained from road calibration. The region of interest includes the specific location coordinates of motor vehicle lanes, non-motor vehicle lanes, and intersections. The specific location coordinates of intersections use a left-handed coordinate system. The 4D millimeter-wave radar is used as the coordinate origin. The point cloud of the 4D millimeter-wave radar is filtered, and point cloud preprocessing is achieved by setting thresholds for the height and speed attributes of the non-ROI area.

3. The method for achieving track stabilization of vehicles queuing at intersections based on 4D millimeter-wave radar according to claim 1, characterized in that: Step 3: Perform DBSCAN clustering on the preprocessed 4D millimeter wave point cloud to obtain the clustering box containing the center coordinate position, length, width, height and average radial velocity. The specific method is: DBSCAN clustering is performed based on the three-dimensional coordinate position and relative radial velocity in the preprocessed millimeter-wave radar point cloud information. A radar point cloud is selected as the core point, and the difference in three-dimensional coordinate position and relative radial velocity between this point and the adjacent radar point cloud is calculated to see if it meets the clustering parameters. If the clustering parameters are met, the adjacent radar points are judged as qualified adjacent points, and the number of adjacent points is calculated. After finding the adjacent points of the radar point cloud, the adjacent points are used as the core points to continue looking for adjacent points. The steps of finding adjacent points are repeated until all adjacent points are found, and a cluster containing n points is obtained. According to the characteristics of millimeter-wave radar point cloud, two sets of clustering parameters are set with 80m of road as the boundary. The specific parameters within 80m are: 80m away: Among them, x core ,y core , z core Indicates the three-dimensional distance of the core point relative to the millimeter wave radar, vr core Indicates the relative radial velocity of the core point relative to the millimeter wave radar, x border ,y border , z border Indicates the three-dimensional distance of the nearby point relative to the millimeter wave radar, vr border Indicates the relative radial velocity of the nearby point relative to the millimeter-wave radar; Select the maximum and minimum values ​​in the x direction in the cluster, that is, x max ,x min And the maximum and minimum values ​​in the y direction, that is, y max ,y min , use these four points to get the cluster bounding box; The center coordinate x of the bounding box cneter =(x max +x min ) / 2,y cneter =(y max +y min ) / 2, the length of the bounding box is length = x max -x min , width = y max -y min , find the maximum and minimum values ​​in the z direction in the cluster frame, that is, z max ,z min , height=z max -z min , add up the relative radial velocities of all radar points in the cluster frame to get vr total , the average radial velocity vr is obtained according to the number of radar points in the cluster box mean =vr total / n.

4. The method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter wave radar according to claim 1, characterized in that: Step 4 associates the generated track with the clustered point cloud cluster. The specific method for generating new tracks from unassociated clusters is as follows: Perform global matching on clusters and tracks, calculate the association scores between clusters and tracks, select a group of clusters with the largest association scores and tracks for association; for clusters without corresponding tracks, create a new track.

5. The method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter wave radar according to claim 4, characterized in that: The association score between clusters and tracks is composed of a weighted combination of distance score, area score, and speed score. The calculation process is: The distance score is calculated according to the following formula: Among them, x c is the distance of the clustering box in the x direction, x t is the distance of the track in the x direction, y c is the distance of the clustering box in the y direction, y t is the distance of the track in the y direction, dist cri is the preset distance threshold; The area score is calculated according to the following formula: Among them, area c is the area of ​​the cluster box, area t is the area of ​​the track, min represents the minimum value, and max represents the maximum value; The speed score is calculated according to the following formula: Among them, VR c is the average radial velocity of the cluster box, vr t is the average radial velocity of the track, vr cri is the preset speed threshold; The association score is calculated according to the following formula: score total =0.4×score distance +0.2×score area +0.4×score velocity 。 6. The method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter wave radar according to claim 1, characterized in that: Step 6.1: When the track enters the queue area, a lane track index is created to store the lane track index number of the current frame. The specific method for mapping the track to the lane is as follows: When the target track enters the queuing area, that is, the coordinate x of the target track meets the preset range gate, then x track –x cross <d cri , where x track is the distance of the track in the x direction of the kth frame, x cross The distance of the intersection in the x direction, d cri For the queueing area threshold, the target has formed a stable track. A new lane track index is created, and the index number of the target track and the corresponding lane number are stored in it. Each lane is stored in the order of the smallest track coordinate x distance from the intersection.

7. The method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter wave radar according to claim 1, characterized in that: Step 6.2 searches for the lane track index of the previous frame and compares it with the lane track index of the current frame. The specific method for querying lost tracks is as follows: According to the lane track index of the kth frame, start searching from the first target in the i-th lane in the track index of the lane in the k+1th frame. If the track target is not found, search for the track target in the lanes i–1 and i+1. If the track target is not found, it is marked as lost, where i is the lane number of the road calibration information.

8. The method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter wave radar according to claim 1, characterized in that: Step 6.3 is as follows: Check the lost target and set a threshold for the number of detection frames and speed of the lost track. If the speed is less than the threshold, it is judged as a false target to reduce the false tracks formed by ghost points in the millimeter wave radar point cloud. Check the distance between the lost target and the stop line to determine whether the first vehicle in the team has left. If the first vehicle in the team has left, check whether there is target movement in the current frame. If there is target movement, do not modify it. Otherwise, keep the target to prevent the first vehicle in the team from being lost. Recall the lost track and add the track attribute value back to the track management.

9. The method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter wave radar according to claim 1, characterized in that: The step 6.4 is specifically as follows: According to the lane track index of the k+1th frame, start searching from the first target in the i-th lane in the track index of the k-th frame. If a new track is found in the i-th lane of the k+1th frame, check whether the i-1 and i+1 lanes of the k-th frame have the same track. If not, the track is considered to be a new track, where i is the lane number of the road calibration information.

10. The method for achieving track stabilization of vehicles queuing at an intersection based on 4D millimeter wave radar according to claim 1, characterized in that: The specific method of performing K-Means clustering on the queuing area point cloud in step 6.6 is: The number of lane tracks is obtained from the lane track index of the current frame. The 4D millimeter-wave radar point cloud of the queuing area is clustered by lane. K-Means clustering is performed on the millimeter-wave radar point cloud of the current lane. The number of clusters is set to the number of lane tracks k. K points are selected from the current lane point cloud as cluster centers, and the Euclidean distances of all points to these k points are calculated, that is, Among them, x c is the distance of the cluster center in the x direction, x i is the distance of the i-th radar point in the x direction, y c is the distance of the cluster center in the y direction, y i is the distance of the i-th radar point in the y direction; Assign the point with the minimum Euclidean distance from the cluster center to the cluster center. After all points are assigned, k clusters are formed. Calculate the mean of each cluster and update the k cluster centers, that is, Among them, x c is the distance of the cluster center in the x direction, x j is the distance of the jth radar point in the cluster in the x direction, y c is the distance of the cluster center in the y direction, and y is the distance of the jth radar point in the cluster in the y direction; Calculate the difference between the new cluster center and the previous cluster center. When the difference is less than 0.0001 or the number of iterations reaches the set number, end the iteration and complete the point cloud clustering of the current lane. Repeat this step until the point clouds of all lanes are clustered.

Images