A clustering method based on traffic millimeter wave radar point cloud data

By improving the DBSCAN algorithm in traffic scenarios and using rectangles to extend neighborhood and inter-frame clustering, the problems of sparseness and target number uncertainty of millimeter-wave radar point cloud data are solved, and the target tracking effect is improved.

CN115372972BActive Publication Date: 2024-11-05XIDIAN UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210939065.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-08-05
Publication Date
2024-11-05
Estimated Expiration
2042-08-05

AI Technical Summary

Technical Problem

In traffic scenarios, the sparsity of millimeter-wave radar point cloud data and uncertainty in target number make it difficult to obtain good clustering effects using the DBSCAN algorithm alone, affecting the subsequent target tracking effect.

Method used

Using a clustering method based on traffic millimeter-wave radar point cloud data, the rectangular neighborhood parameters and velocity are used by changing the original DBSCAN circular extended neighborhood into a rectangular extended neighborhood, and combining intra-clustering and inter-clustering. Difference threshold, set Minpts to 1 and 2 to retain or remove noise points to ensure accurate clustering of target points.

Benefits of technology

It effectively solves the problem of small targets without trajectory at long distances and large targets in close distances, and improves the clustering quality of point cloud data and target tracking effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115372972B_ABST
    Figure CN115372972B_ABST
Patent Text Reader

Abstract

The present invention discloses a clustering method based on traffic millimeter wave radar point cloud data, the method comprising the following steps: obtaining point cloud data of the current frame; performing intra-frame DBSCAN processing on the current frame data; obtaining the clustering result of the previous frame; performing inter-frame clustering using the clustering result of the current frame and the clustering result of the previous frame; outputting the clustering result and performing point trace condensation. The present invention solves two problems: small targets have no tracks at long distances and large targets are split at close distances, improves the quality of point cloud data clustering, and thus improves the quality of target tracking.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of radar signal processing, and in particular to a clustering method based on traffic millimeter-wave radar point cloud data. Background Art

[0002] In traffic scenarios, millimeter-wave radar tracking of moving targets is generally divided into two parts: clustering and tracking. The output points after point cloud data clustering serve as the input of the subsequent target tracking module. Therefore, the quality of point cloud target clustering has a huge impact on the subsequent tracking effect.

[0003] Currently, in traffic road scenes, millimeter-wave radar point cloud data presents two characteristics: sparsity and uncertainty in the number of targets. This determines that density-based clustering algorithms (such as DBSCAN) are more suitable for this scenario. However, in actual road scenes, the characteristics of point cloud data are as follows: small targets and long-distance targets have only a small number of points or even only one point, and large targets have a large number of reflection points, but the distribution range is very scattered. Moreover, for the same target, due to the different RCS at different locations, the point cloud density at different locations is also different. Therefore, if only intra-frame clustering is performed and the parameter Minpts is set to 1, noise points will be judged as targets, and if the parameter Minpts value is set to a value greater than or equal to 2, targets with only one reflection point will be removed as noise points, which will cause the subsequent tracking of targets to deteriorate.

[0004] In summary, it is difficult to obtain good results by using the DBSCAN algorithm alone to perform intra-frame clustering of point cloud data. How to improve the point cloud clustering effect and provide better point data for the tracking module is an urgent problem to be solved in this field. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention aims to provide a clustering method based on traffic millimeter-wave radar point cloud data, which realizes:

[0006] 1. In view of the rectangular shape of vehicle point cloud data in traffic scenarios, the original DBSCAN circular extended neighborhood is improved to a rectangular extended neighborhood, which can more accurately cluster the data of the same target point into a cluster.

[0007] 2. In order to solve the problems of small targets, long-distance targets and large targets in traffic scenes, the present invention first performs intra-frame clustering and then inter-frame clustering, which can effectively solve the two difficult problems of small targets having no tracks at a long distance and large targets being split at a close distance, thereby improving the quality of point cloud data clustering and thus improving the quality of target tracking.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A clustering method based on traffic millimeter wave radar point cloud data, the method comprising the following steps:

[0010] S1 obtains the point cloud data of the current frame;

[0011] S2 performs intra-frame DBSCAN processing on the current frame data;

[0012] S3 obtains the clustering result of the previous frame;

[0013] S4 performs inter-frame clustering using the clustering result of the current frame and the clustering result of the previous frame;

[0014] S5 outputs the clustering results and performs point agglomeration;

[0015] The step S4 further includes:

[0016] S4.1 Obtain the cluster centroid data information of the previous frame data S′={S1′, S2′, S3′…, S K ′}, perform inter-frame clustering with the cluster centroid data of the current frame;

[0017] S4.2 Determine the parameters in the inter-frame clustering: the rectangular neighborhood parameters of the inter-frame clustering longitudinal neighborhood eps_rx2, the inter-frame clustering lateral neighborhood eps_ry2, the inter-frame clustering speed neighborhood eps_v2, and the minimum number of points contained in each cluster Minpts2 = 2, so that the noise points retained during the intra-frame clustering can be removed, and only the point trace data with the connection between the previous and next frames can be retained;

[0018] S4.3 condenses the sample points of the same frame and cluster respectively to obtain the final centroid information of the current frame cluster.

[0019] It should be noted that, in the step S1, it also includes:

[0020] S1.1 Get the original target point data of the current frame point cloud {P i}, i = 1, 2, 3, ..., N, where N is the number of original target points, and the information of each target point includes P i = {r i ,v i ,θ i, F i},r i is the distance information of the target, v i is the target speed information, θ i is the angle information of the target, F i The frame number information of the target;

[0021] S1.2 The original target point data {P i}Transform to rectangular coordinate system to get P i′={r i , x i ,y i ,θ i , F i}, where x i and i They are the position information of the original point trace in the rectangular coordinate system;

[0022] x i =r i *cosθ i ,y i =r i *sinθ i .

[0023] It should be noted that, in the step S2, it also includes determining the parameters in the intra-frame clustering: the rectangular neighborhood parameters longitudinal distance eps_rx1, lateral distance eps_ry1 and speed difference threshold eps_v1, and the minimum number of points contained in each cluster Minpts1=1; wherein, by observing the speed in the echo data, it is found that the speed difference of the same target is very small, so the speed parameter is introduced, and the speed difference threshold eps_v1 is set to better cluster the echo target;

[0024] Initialize the number of clusters k1 = 0, and the unvisited sample set Γ1 = P i ′, where i = 1, 2, 3, ..., N, cluster partition set

[0025] The point data are processed by DBSCAN algorithm and divided into clusters to obtain the cluster partition set C = {C1, C2, ..., C k}, according to the sample points of each cluster, find the centroid of the cluster, and use the centroid information S k Instead of the point cloud information of each cluster, that is, summing and averaging each sample point p belonging to the cluster, we get Then the cluster centroid information set S = {S1, S2, ..., S k}.

[0026] It should be noted that, according to the rectangular distribution of the echo data of vehicles in traffic, the neighborhood is expanded to avoid the vehicles in adjacent lanes from being clustered into one target or the over-long vehicles from being clustered into two targets as much as possible.

[0027] It should be noted that the step S4.3 also includes:

[0028] After inter-frame clustering, each cluster is divided into sets C' = {C1', C2', ..., C m The sample points of each cluster in ′} are condensed, and the sample points of the current frame number are selected in each cluster for point condensation. The centroid information after condensation is: S j= {r j , x j ,y j , v j ,θ j , F now}, j = 1, 2, ... mm is the number of partition clusters, F now is the current frame number; where:

[0029]

[0030] In the formula, a represents the point number in each cluster, and n represents the total number of points in each cluster, so a = 1, 2, ..., n.

[0031] The beneficial effect of the present invention is that the present invention first sets the DBSCAN parameter Minpts to 1 to ensure that when the small target is at a long distance, there is only one echo target point and it will not be removed as noise. However, this will cause a single noise point to be treated as a target point and the point cloud data to be relatively scattered when the large target is at a close distance, thereby causing the large target to split. Therefore, the clustered data of the previous frame is taken out and inter-frame clustering is performed with the current frame data, and the parameter Minpts is set to 2 to remove noise points. Moreover, after the connection between the upper and lower frames, if the large target suddenly splits at a close distance, that is, a new target appears, but the previous frame does not have clustering information of the target, the split point traces can be treated as noise and removed. In this way, the two difficult problems of no track for small targets at a long distance and target splitting for large targets at a close distance are solved, and the quality of point cloud data clustering is improved, thereby improving the quality of target tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flow chart of a clustering method based on millimeter wave radar point cloud data in an embodiment;

[0033] Figure 2 is an image of the real environment;

[0034] Figure 3 For intra-frame clustering, the clustering result when the parameter minpts is set to 1;

[0035] Figure 4 For inter-frame clustering, the clustering result when the parameter minpts is set to 2;

[0036] Figure 5 It is the clustering result obtained by combining intra-frame clustering and inter-frame clustering. DETAILED DESCRIPTION

[0037] The present invention will be further described below. It should be noted that the following embodiments are based on the present technical solution and provide detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to these embodiments.

[0038] The present invention provides a clustering method based on traffic millimeter wave radar point cloud data, and the method comprises the following steps:

[0039] S1 obtains the point cloud data of the current frame;

[0040] S2 performs intra-frame DBSCAN processing on the current frame data;

[0041] S3 obtains the clustering result of the previous frame;

[0042] S4 performs inter-frame clustering using the clustering result of the current frame and the clustering result of the previous frame;

[0043] S5 outputs the clustering results and performs point agglomeration;

[0044] The step S4 further includes:

[0045] S4.1 Obtain the cluster centroid data information of the previous frame data S′={S1′, S2′, S3′…, S K ′}, perform inter-frame clustering with the cluster centroid data of the current frame;

[0046] S4.2 Determine the parameters in the inter-frame clustering: the rectangular neighborhood parameters of the inter-frame clustering longitudinal neighborhood eps_rx2, the inter-frame clustering lateral neighborhood eps_ry2, the inter-frame clustering speed neighborhood eps_v2, and the minimum number of points contained in each cluster Minpts2 = 2, so that the noise points retained during the intra-frame clustering can be removed, and only the point trace data with the connection between the previous and next frames can be retained;

[0047] S4.3 condenses the sample points of the same frame and cluster respectively to obtain the final centroid information of the current frame cluster.

[0048] Furthermore, the present invention further comprises, in step S1:

[0049] S1.1 Get the original target point data of the current frame point cloud {P i}, i = 1, 2, 3, ..., N, where N is the number of original target points, and the information of each target point includes P i = {r i ,v i ,θ i, F i},r i is the distance information of the target, v i is the target speed information, θ i is the angle information of the target, Fi The frame number information of the target;

[0050] S1.2 The original target point data {P i}Transform to rectangular coordinate system to get P i ′={r i , x i ,y i ,θ i , F i}, where x i and i They are the position information of the original point trace in the rectangular coordinate system;

[0051] x i =r i *cosθ i ,y i =r i *sinθ i .

[0052] Furthermore, in the step S2, the present invention further includes determining parameters in the intra-frame clustering: rectangular neighborhood parameters longitudinal distance eps_rx1, lateral distance eps_ry1 and speed difference threshold eps_v1, and the minimum number of points contained in each cluster Minpts1=1; wherein, the speed in the echo data is observed, and it is found that the speed difference of the same target is very small, so the speed parameter is introduced, and the speed difference threshold eps_v1 is set to better cluster the echo target;

[0053] Initialize the number of clusters k1 = 0, and the unvisited sample set Γ1 = P i ′, where i = 1, 2, 3, ..., N, cluster partition set

[0054] The point data are processed by DBSCAN algorithm and divided into clusters to obtain the cluster partition set C = {C1, C2, ..., C k}, according to the sample points of each cluster, find the centroid of the cluster, and replace the point cloud information of each cluster with the centroid information Si, that is, sum and average each sample point p belonging to the cluster to get Then the cluster centroid information set S = {S1, S2, ..., S k}.

[0055] It should be noted that, according to the rectangular distribution of the echo data of vehicles in traffic, the neighborhood is expanded to avoid the vehicles in adjacent lanes from being clustered into one target or the over-long vehicles from being clustered into two targets as much as possible.

[0056] Furthermore, the present invention further comprises, in step S4.3:

[0057] After clustering between frames, each cluster is divided into sets C′={C1′, C2′, …, C m The sample points of each cluster in ′} are condensed, and the sample points of the current frame number are selected in each cluster for point condensation. The centroid information after condensation is: S i = {r i ,x i ,y i ,v i ,θ i ,F now}, i = 1, 2, 3, ..., kk is the number of clusters, F now is the current frame number; where:

[0058]

[0059]

[0060] In the formula, a represents the point number in each cluster, and n represents the total number of points in each cluster, so a = 1, 2, ..., n.

[0061] Example 1

[0062] The effectiveness of the present invention will be further described below by way of Example 1.

[0063] Step 1: Get the current frame data.

[0064] Step 1-1: Get the original target point data of the current frame point cloud {P i}, i = 1, 2, 3, ..., N, where N is the number of original target points, and the information of each target point includes P i ={ri,v i ,θ i ,F i},r i is the distance information of the target, v i is the target speed information, θ i is the angle information of the target, F i The frame number information of the target.

[0065] Step 1-2: Get the original target point data of the current frame point cloud {P i}Transform to rectangular coordinate system to get P i ′={r i ,x i ,y i ,v i ,θ i ,F i}, where x i and i They are the position information of the original points in the rectangular coordinate system.

[0066] x i =r i *cosθ i ,y i =r i *sinθ i

[0067] Step 2: perform intra-frame clustering and point condensation.

[0068] Step 2-1: Determine the parameters in the intra-frame clustering: rectangular neighborhood parameters longitudinal distance eps_rx1, lateral distance eps_ry1, velocity difference threshold eps_v1, and the minimum number of points contained in each cluster Minpts1=1, so as to avoid the target being judged as a noise point when there is only one echo data.

[0069] Step 2-2: Initialize the number of clusters k1 = 0 and the set of unvisited samples Γ1 = P i ′, where i = 1, 2, 3, ..., N, cluster partition set

[0070] Step 2-3: The point data are processed by the DBSCAN algorithm and divided into clusters to obtain the cluster partition set C = {C1, C2, ..., C k}, according to the sample points of each cluster, find the centroid of the cluster, and replace the point cloud information of each cluster with the centroid information Si, that is, sum and average each sample point p belonging to the cluster to get Then the cluster centroid information set S = {S1, S2, ..., S k}.

[0071] Step 3, obtain the clustered centroid data of the previous frame and perform inter-frame clustering with the clustering result of the current frame.

[0072] Step 3-1: Obtain the cluster centroid data information S′={S1′, S2′, S3′…, S K ′}, and perform inter-frame clustering with the cluster centroid data of the current frame.

[0073] Step 3-2: Determine the parameters in inter-frame clustering: rectangular neighborhood parameters - longitudinal neighborhood eps_rx2 of secondary clustering, lateral neighborhood eps_ry2 of secondary clustering, velocity neighborhood eps_v2 of secondary clustering, and the minimum number of points contained in each cluster Minpts2=2. This can remove the noise points retained during intra-frame clustering and only retain the point data that are connected to the previous and next frames.

[0074] Step 4: Perform point condensation on the sample points in the same frame and cluster to obtain the final centroid information of the current frame cluster. mThe sample points of each cluster in ′} are condensed, and the sample points of the current frame number are selected in each cluster for point condensation. The centroid information after condensation is: S j = {r j , x j ,y j , v j ,θ j , F now}, j = 1, 2, ... mm is the number of partition clusters, F now is the current frame number.

[0075]

[0076] In the formula, a represents the point number in each cluster, and n represents the total number of points in each cluster, so a = 1, 2, ..., n.

[0077] like Figures 2 to 5 In the specific embodiment shown, Figure 2 is an image of the real environment, Figure 3 , Figure 4 These are the clustering effects when only intra-frame clustering is performed and the parameter minpts is set to 1 and 2. Figure 5 This is the effect of the embodiment of the present application. It should be noted that the data used in this application example is 500 frames of continuous radar point cloud data, and the maneuvering targets are clustered, so the static point cloud has been screened out in the early data processing. The present invention can accurately cluster the targets in each frame, and when only intra-frame clustering is performed, the parameter minpts is set to 1, which causes the target splitting phenomenon of vehicle No. 4. When the parameter minpts is set to 2, the two small target vehicles 1 and 3 will be removed as noise points at a long distance. By comparison, it is obvious that the present invention is better than only intra-frame clustering.

[0078] Although DBSCAN can cluster point cloud data of any shape, in traffic scenarios, the point cloud data of motor vehicles is generally rectangular. The present invention fully considers the distribution state of vehicle point cloud data and uses rectangular neighborhoods to cluster point cloud data, which can avoid adjacent small target vehicles being clustered into one cluster or large target vehicles being clustered into multiple clusters.

[0079] For those skilled in the art, various corresponding changes and modifications can be made according to the above technical solutions and concepts, and all of these changes and modifications should be included in the protection scope of the claims of the present invention.

Claims

1. A clustering method based on traffic millimeter wave radar point cloud data, It is characterized in that The method comprises the following steps: S1 obtains the point cloud data of the current frame; S2 performs intra-frame DBSCAN processing on the current frame data; S3 obtains the clustering result of the previous frame; S4 performs inter-frame clustering using the clustering result of the current frame and the clustering result of the previous frame; S5 outputs the clustering results and performs point agglomeration; In the step S4, it also includes: S4.1 Obtain the cluster centroid data information S′ of the previous frame data = {S 1 ′,S 2 ′,S 3 ′…,S K ′}, perform inter-frame clustering with the cluster centroid data of the current frame; S4.2 Determine the parameters in the inter-frame clustering: the rectangular neighborhood parameter inter-frame clustering longitudinal neighborhood eps_rx2, the inter-frame clustering lateral neighborhood eps_ry2, the inter-frame clustering speed neighborhood eps_v2, and the minimum number of points contained in each cluster Minpts2 = 2, so that the noise points retained during the intra-frame clustering can be removed, and only the point trace data with the connection between the previous and next frames can be retained; S4.3 condenses the sample points of the same frame and cluster respectively to obtain the final centroid information of the current frame cluster.

2. The clustering method based on traffic millimeter wave radar point cloud data according to claim 1, It is characterized in that In the step S1, it also includes: S1.1 Get the original target point data of the current frame point cloud {P i }, i = 1, 2, 3, ..., N, where N is the number of original target points, and the information of each target point includes P i = {r i ,v i ,θ i, F i },r i is the distance information of the target, v i is the target speed information, θ i is the angle information of the target, F i The frame number information of the target; S1.2 The original target point data of the current frame point cloud obtained {P i }Transform to rectangular coordinate system to get P i '={r i ,v i ,x i ,y i ,θ i ,F i }, where x i and i They are the position information of the original points in the rectangular coordinate system: x i =r i *cosθ i ,y i =r i *sinθ i 。 3. The clustering method based on traffic millimeter wave radar point cloud data according to claim 2, It is characterized in that In the step S2, it also includes determining parameters in intra-frame clustering: rectangular neighborhood parameters longitudinal distance eps_rx1, lateral distance eps_ry1 and speed difference threshold eps_v1, and the minimum number of points contained in each cluster Minpts1=1; wherein, by observing the speed in the echo data, it is found that the speed difference of the same target is very small, so the speed parameter is introduced, and the speed difference threshold eps_v1 is set to better cluster the echo target; Initialize the number of clusters k1 = 0, and the unvisited sample set Γ1 = P′ i , where i = 1, 2, 3, ..., N, cluster partition set The point data are processed by DBSCAN algorithm and divided into clusters to obtain the cluster partition set C = {C 1 , C 2 , …, C k }, according to the sample points of each cluster, find the centroid of the cluster, and use the centroid information S k Instead of the point cloud information of each cluster, that is, summing and averaging each sample point p belonging to the cluster, we get Then the cluster centroid information set S = {S 1 ,S 2 ,…,S k }.

4. The clustering method based on traffic millimeter wave radar point cloud data according to claim 3, It is characterized in that According to the rectangular distribution of echo data of vehicles in traffic, the neighborhood is expanded to avoid the vehicles in adjacent lanes from being clustered into one target or the over-long vehicles from being clustered into two targets as much as possible.

5. The clustering method based on traffic millimeter wave radar point cloud data according to claim 1, It is characterized in that The step S4.3 further includes: dividing each cluster after inter-frame clustering into sets C'={C 1 ′,C 2 ′,…,C m The sample points of each cluster in ′} are condensed, and the sample points of the current frame number are selected in each cluster for point condensation. The centroid information after condensation is: S j = {r j , x j ,y j , v j ,θ j , F now }, j = 1, 2, ... mm is the number of partition clusters, F now is the current frame number; where: Where a represents the point number in each cluster, and n represents the total number of points in each cluster, so a = 1, 2, ..., n.