A method for clustering millimeter-wave radar point cloud targets based on KDE-DBSCAN

Abstract

The present invention discloses a method for clustering millimeter-wave radar point cloud targets based on KDE-DBSCAN, comprising the following steps: S1. For each point in the millimeter-wave radar point cloud data set D, calculate the corresponding associated window width W and the associated adjacent points RN of this point according to the k value corresponding to the region where this point is located; S2. According to the associated window width W and the set of associated adjacent points RN of each point, preliminarily divide the radar point cloud data set D into different pseudo-clusters FC; S3. Count the number of points in each pseudo-cluster FC generated in step S2 i , and delete the pseudo-clusters with the number of points less than λ; S4. Determine the optimal associated radius ε and the minimum number of associated points minP of each pseudo-cluster; S5. From the optimal associated radius ε and the minimum number of associated points minP of each pseudo-cluster obtained in step S4, then use the DBSCAN algorithm to cluster each pseudo-cluster respectively. The method of the present invention can adapt to the point cloud distribution characteristics of the millimeter-wave radar, thereby improving the clustering quality of the millimeter-wave radar targets.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud target segmentation, clustering and recognition, and in particular to a millimeter wave radar point cloud target clustering method based on KDE-DBSCAN. Background Art

[0002] Millimeter-wave radar has the functions of measuring distance, speed and angle of targets, and has good robustness in many working conditions, so it is increasingly used in autonomous driving. In the application of millimeter-wave radar, radar target point cloud clustering is a very important part. As the resolution of millimeter-wave radar increases, the amount of data reflected by the radar from the same target also increases. Therefore, it is necessary to use a suitable clustering algorithm to cluster the target point cloud of millimeter-wave radar.

[0003] At present, the commonly used clustering algorithms for millimeter-wave radars are K-means clustering, DBSCAN clustering, etc. However, these algorithms work better when processing data sets with uniform data density. However, the data density of the millimeter-wave radar target point cloud is uneven. Since the angular resolution of the millimeter-wave radar in a certain direction is fixed and the angular resolution in different directions is different, the density of the target point cloud is closely related to the distance and angle of the target. Generally, the closer the distance and the smaller the angle, the greater the density of the target point cloud. The farther the distance and the larger the angle, the sparser the target point cloud. At this time, it is difficult for the above commonly used clustering algorithms to perform well. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a millimeter-wave radar point cloud target clustering method based on KDE-DBSCAN, which can adapt to the point cloud distribution characteristics of millimeter-wave radar, thereby improving the clustering quality of millimeter-wave radar targets.

[0005] In order to solve the above technical problems, the technical solution of the present invention is: a millimeter wave radar point cloud target clustering method based on KDE-DBSCAN, comprising the following steps:

[0006] S1. For each point in the millimeter-wave radar point cloud data set D, calculate the corresponding correlation window width W and the associated adjacent points RN of the point according to the k value corresponding to the area where the point is located;

[0007] S2, based on the association window width W of each point and the associated adjacent point set RN, the radar point cloud data set D is preliminarily divided into different pseudo clusters FC;

[0008] S3, count the FC of each pseudo cluster generated in step S2 i The number of points in, delete the pseudo clusters with less than λ points;

[0009] S4, determining the optimal correlation radius ε and the minimum number of correlation points minP of each pseudo cluster;

[0010] S5. Obtain the optimal correlation radius ε and the minimum number of correlation points minP of each pseudo cluster from step S4, and then use the DBSCAN algorithm to cluster each pseudo cluster respectively.

[0011] As a preferred technical solution, regarding step S1, the details are as follows:

[0012] S1-1, divided into N regions according to radial distance, N regions are denoted as r 0 , r 1 ,……,r N-1 ;

[0013] S1-2, traverse each point x in the point cloud data set D, and find the k points closest to point x according to the area where x is located. ri adjacent points, generate the Euclidean distance set D nn (x) = {d j =d(x,x j )|j=[1,......,k ri ]}, where ri = 0, 1, ..., N-1, and d j Sort in ascending order;

[0014] S1-3, select the kernel density function K(x), use K(x) and the Euclidean distance set D in step S1-2 nn (x) to generate the distance probability density distribution function where h is the bandwidth of the kernel function K(x), i = 1, 2, ..., k ri ;

[0015] S1-4. Based on the generated distance probability density distribution function Find the first maximum point of the function {(x j ,f(x j ))|1<j≤h,h≤k ri}, then x j＝[1，...，h] The associated neighboring point RN belongs to this point, and the associated window width W of this point is equal to d h .

[0016] As a preferred technical solution, regarding step S2, the details are as follows:

[0017] S2-1, arranging the associated window width W of each point in step S1 in ascending order;

[0018] S2-2, select the point with the smallest correlation window width W, traverse each point in the data set D, if the point has not been attributed to any pseudo cluster, then recursively add the associated adjacent points belonging to the point x to the pseudo cluster FC i middle;

[0019] S2-3. Remove the pseudo cluster FC from the radar dataset D. i point, and i=i+1;

[0020] S2-4. Repeat steps S2-2 to S2-3 until all points in the data set D are traversed.

[0021] As a preferred technical solution, regarding step S4, the details are as follows:

[0022] S4-1. For a given pseudo cluster FC i ={x j |j≤n i}, the associated window width W of all points constitutes the set {w m |m≤n i}, record the pseudo cluster inner point x j At a radius of w m The number of adjacent points in the range is m j , then define the function

[0023] Among them, RN j is a pseudo-cluster point x j The number of associated adjacent points, search for w that minimizes the function f m , then this w m is the optimal correlation radius of this pseudo-cluster;

[0024] S4-2. For a given pseudo cluster FC i ={x j |j≤n i}, can be obtained by formula

[0025] To determine the minimum number of associated points threshold minP,

[0026] Where k = 1, 2, ..., n'; n' is the number of pseudo clusters, n k FC is a pseudo cluster k The number of points within, totalVolume k is the volume of the smallest hyperrectangle that contains all the points in the pseudo-cluster. The hyperrectangle can be determined by the maximum and minimum values ​​of each dimension of the data; εVolume k It can be determined based on the optimal correlation radius ε in step 4. For 1-D data, εVolume k is 2*ε; for 2-D data, the value is π*ε2 , for 3-D data, the value is For higher dimensional data, the value is the volume of a hypercube with side length 2*ε.

[0027] Due to the adoption of the above technical solution, the beneficial effect of the present invention is that the method of the present invention can adapt to the point cloud distribution characteristics of the millimeter wave radar, thereby improving the clustering quality of the millimeter wave radar target. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The following drawings are intended only to illustrate and explain the present invention, and are not intended to limit the scope of the present invention.

[0029] Figure 1 is a flow chart of an embodiment of the present invention; DETAILED DESCRIPTION

[0030] The present invention is further described below in conjunction with the accompanying drawings and examples. In the following detailed description, certain exemplary embodiments of the present invention are described only by way of illustration. Needless to say, those of ordinary skill in the art will recognize that the described embodiments may be modified in various ways without departing from the spirit and scope of the present invention. Therefore, the drawings and description are illustrative in nature and are not intended to limit the scope of protection of the claims.

[0031] like Figure 1 As shown, a millimeter wave radar point cloud target clustering method based on KDE-DBSCAN includes the following steps:

[0032] S1. For each point in the millimeter-wave radar point cloud data set D, calculate the corresponding correlation window width W and the associated adjacent points RN of the point according to the k value corresponding to the area where the point is located;

[0033] Regarding step S1, the details are as follows:

[0034] S1-1, divided into N regions according to radial distance, N regions are denoted as r 0 , r 1 ,……,r N-1 ;

[0035] S1-2, traverse each point x in the point cloud data set D, and find the k points closest to point x according to the area where x is located. ri adjacent points, generate the Euclidean distance set D nn (x) = {d j =d(x,x j )|j=[1,......,k ri ]}, where ri = 0, 1, ..., N-1, and d j Sort in ascending order;

[0036] S1-3, select the kernel density function K(x), use K(x) and the Euclidean distance set D in step S1-2 nn (x) to generate the distance probability density distribution function where h is the bandwidth of the kernel function K(x), i = 1, 2, ..., k ri ;

[0037] S1-4. Based on the generated distance probability density distribution function Find the first maximum point of the function {(x j ,f(x j ))|1<j≤h,h≤k ri}, then x j＝[1，...，h] The associated neighboring point RN belongs to this point, and the associated window width W of this point is equal to d h .

[0038] S2, based on the association window width W of each point and the associated adjacent point set RN, the radar point cloud data set D is preliminarily divided into different pseudo clusters FC;

[0039] Regarding step S2, the details are as follows:

[0040] S2-1, arranging the associated window width W of each point in step S1 in ascending order;

[0041] S2-2, select the point with the smallest correlation window width W, traverse each point in the data set D, if the point has not been attributed to any pseudo cluster, then recursively add the associated adjacent points belonging to the point x to the pseudo cluster FC i middle;

[0042] S2-3. Remove the pseudo cluster FC from the radar dataset D. i point, and i=i+1;

[0043] S2-4. Repeat steps S2-2 to S2-3 until all points in the data set D are traversed.

[0044] S3, count the FC of each pseudo cluster generated in step S2 i The number of points in, delete the pseudo clusters with less than λ points;

[0045] S4, determining the optimal correlation radius ε and the minimum number of correlation points minP of each pseudo cluster;

[0046] Regarding step S4, the details are as follows:

[0047] S4-1. For a given pseudo cluster FC i ={x j |j≤n i}, the associated window width W of all points constitutes the set {w m |m≤n i}, record the pseudo cluster inner point x j At a radius of w m The number of adjacent points in the range is m j , then define the function

[0048] Among them, RN j is a pseudo-cluster point x j The number of associated adjacent points, search for w that minimizes the function f m , then this w m is the optimal correlation radius of this pseudo-cluster;

[0049] S4-2. For a given pseudo cluster FC i ={x j |j≤n i}, can be obtained by formula

[0050] To determine the minimum number of associated points threshold minP,

[0051] Where k = 1, 2, ..., n'; n' is the number of pseudo clusters, n k FC is a pseudo cluster k The number of points within, totalVolume k is the volume of the smallest hyperrectangle that contains all the points in the pseudo-cluster. The hyperrectangle can be determined by the maximum and minimum values ​​of each dimension of the data; εVolume k It can be determined based on the optimal correlation radius ε in step 4. For 1-D data, εVolume k is 2*ε; for 2-D data, the value is π*ε 2 , for 3-D data, the value is For higher dimensional data, the value is the volume of a hypercube with side length 2*ε.

[0052] S5. Obtain the optimal correlation radius ε and the minimum number of correlation points minP of each pseudo cluster from step S4, and then use the DBSCAN algorithm to cluster each pseudo cluster respectively.

[0053] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for clustering millimeter-wave radar point cloud targets based on KDE-DBSCAN, characterized in that, it includes the following steps: S1. For each point in the millimeter-wave radar point cloud data set D, calculate the corresponding associated window width W and the associated neighboring points RN of this point according to the k value corresponding to the area where this point is located. Regarding step S1, specifically as follows: S1-1. Divide it into N regions according to the radial distance, and the N regions are respectively denoted as r 0 , r 1 , ……, r N-1 ; S1-2. Traverse each point x in the point cloud dataset D, and find the k nearest neighboring points to point x according to the region where x is located, generating an Euclidean distance set D ri (x) = {d nn j = d(x, x j )|j = [1, ……, k ri ]}, where ri = 0, 1, ……, N - 1, and d j is sorted in ascending order; S1-3. Select the kernel density function K(x), and use K(x) and the set D of Euclidean distances in step S1-2 nn (x) to generate the distance probability density distribution function where h is the bandwidth of the kernel function K(x), and i = 1, 2,..., k ri ; S1-4. According to the generated distance probability density distribution function Find the first maximum point of the function \(\{(x j , f(x j ))|1 < j ≤ h, h ≤ k ri \}, then \(x j＝[1,…,h] \) belongs to the associated adjacent point \(RN\) of this point, and the associated window width \(W\) of this point is equal to \(d h ;\ S2. According to the associated window width W and the set of associated neighboring points RN of each point, preliminarily divide the radar point cloud data set D into different pseudo-clusters FC. S3. Count the number of points in each pseudo-cluster FC generated in step S2, and delete the pseudo-clusters with the number of points less than λ; i ​ S4. Determine the optimal associated radius ε and the minimum number of associated points threshold minP for each pseudo-cluster. S5. Obtain the optimal associated radius ε and the minimum number of associated points threshold minP for each pseudo-cluster from step S4, and then use the DBSCAN algorithm to cluster each pseudo-cluster respectively.

2. The method for clustering millimeter-wave radar point cloud targets based on KDE-DBSCAN according to claim 1, characterized in that, regarding step S2, specifically as follows: S2-1. Arrange the associated window widths W of each point in step S1 in ascending order. S2-2. Select the point with the smallest associated window width W, and traverse each point in the dataset D. If the point has not been assigned to any pseudo-cluster yet, recursively add the associated adjacent points belonging to this point x to the pseudo-cluster FC i ; S2-3. Remove the points belonging to the pseudo-cluster FC from the radar data set D, and set i = i + 1; i ; S2-4. Repeat steps S2-2 to S2-3 until all points in the data set D have been traversed.

3. The method for clustering millimeter-wave radar point cloud targets based on KDE-DBSCAN according to claim 1 or 2, characterized in that, regarding step S4, specifically as follows: S4-1. For a given pseudo-cluster FC i ={x j | j ≤ n i}, the set of the associated window widths W of all points forms the set {w m | m ≤ n i}. Denote the number of adjacent points of the point x j within the radius of w m as m j . Then, define the function Among them, RN j is the number of associated adjacent points of the pseudo-cluster interior point x j , search for w that minimizes the function f m , then this w m is the optimal associated radius of this pseudo-cluster; S4-2. For a given pseudovariety FC i = {x j | j ≤ n i}, it can be obtained through the formula to determine the minimum number of associated point threshold minP where k = 1, 2, ……, n′; n′ is the number of pseudo-clusters, and n k is the number of points in the pseudo-cluster FC k , and totalVolume k is the volume of the minimum hyper-rectangle that contains all the points in the pseudo-cluster. The hyper-rectangle can be determined by the maximum and minimum values of each dimension of the data; εVolume k can be determined according to the optimal association radius ε in step 4. For 1-D data, εVolume k is 2*ε; for 2-D data, the value is π*ε 2 , and for 3-D data, the value is For higher-dimensional data, the value is the volume of a hypercube with side length 2*ε.

Images