4D millimeter wave radar clustering method and storage medium based on DBSCAN

Abstract

The present invention discloses a 4D millimeter wave radar clustering method and storage medium based on DBSCAN, comprising the following steps: S1, obtaining vehicle signals, lane line signals and 4D millimeter wave radar point cloud signals, and parsing and preprocessing the signals; S2, calculating the local search radius of each point in the point cloud signal, and sorting the calculated local search radius of each point; S3, clustering the point cloud signal based on the DBSCAN algorithm. The present invention is based on the traditional DBSCAN algorithm, in which data preprocessing based on other sensors and the physical properties of each point is added, a local search radius ε is determined for each point cloud according to the size of its physical quantity, all original detections are sorted and clustered, and the number of clustering result points is dynamically judged whether it is a single-frame observation target. Comparing the clustering results of the present invention with the traditional DBSCAN clustering method, it can be found that the clustering effect of the present algorithm is better.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobile intelligent driving, and more specifically relates to a 4D millimeter-wave radar clustering method and a storage medium based on DBSCAN. Background Art

[0002] In the past two years, 4D millimeter-wave radars with multiple chips cascaded and more antenna sending and receiving channels have been widely used in the field of autonomous driving. Compared with ordinary millimeter-wave radars, although their detection performance has been greatly improved and even comparable to lidar, their detection principles and data processing procedures are similar to traditional radars. The detected point clouds need to be clustered before being transmitted to the back end for algorithm operations such as Kalman filtering.

[0003] The angular resolution of the millimeter-wave radar is fixed, which determines that the density of the point cloud reflected by the target is inversely proportional to its distance. When the target is close, the point cloud density is large; when the target is far, the point cloud density is small. The point cloud density detected by the millimeter-wave radar is unevenly distributed, which is manifested as sparse point clouds in the distance and dense point clouds in the near distance. The traditional DBSCAN clustering algorithm is not ideal when dealing with uneven density distribution. When a smaller ε is selected based on a high-density point set, the low-density points will be split into many different clusters; when a higher ε is selected based on a low-density point set, many clusters in the high-density points will be merged into one cluster. Therefore, the traditional DBSCAN algorithm will always have a poor effect on radar point cloud clustering under a certain distance. No matter how the global parameters are selected, it will cause missed detection and false detection of radar data.

[0004] my country's patent 201910661302.6 discloses a millimeter-wave radar multi-target tracking method, which uses a millimeter-wave radar to collect point cloud data, clusters the point cloud data to distinguish the echo signals of different targets, and estimates the state information of the observed multiple targets based on the clustering results, thereby achieving multi-target tracking. This invention improves the DBSCAN clustering algorithm based on the characteristics of millimeter-wave radar, improves the accuracy of estimating the number of targets and target states, and uses Kalman filtering and data association algorithms to achieve prediction and tracking of multiple target trajectories in a more complex environment. It completes the algorithm from point cloud clustering to multi-target tracking. From the details of its implementation, it uses the basic DBSCAN algorithm, that is, it uses the global ε, so it cannot avoid the clustering effect meeting the detection characteristics of millimeter waves.

[0005] my country's patent 202110697622.4 discloses a traffic target recognition method based on the DBSCAN algorithm, which belongs to the field of data processing technology. First, the millimeter-wave radar is used to detect the target to be measured in a continuous time period to obtain different position information of the target to be measured; then the position information is used as point cloud data, and the point cloud data is clustered using the DBSCAN clustering algorithm to obtain each cluster; then, the number of scattering points in the cluster is used to identify and divide the target type, and after obtaining the target type corresponding to each cluster, the number of target types is counted, and finally the identification and counting of traffic targets in a comprehensive traffic environment are completed. This invention improves the accuracy of target recognition, and the recognition process is simple and efficient. It also uses the millimeter-wave radar DBSCAN algorithm and completes the correlation tracking between multiple frames of detection data to count the total number of tracked targets. It still uses global parameters, so the clustering results are still inevitably not in line with the measurement characteristics of the millimeter-wave radar. Summary of the invention

[0006] To solve the above problems, the present invention provides a 4D millimeter-wave radar clustering method and storage medium based on DBSCAN, so that the entire clustering algorithm is more in line with the characteristics of millimeter-wave radar.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows: a 4D millimeter wave radar clustering method based on DBSCAN, comprising the following steps:

[0008] S1. Obtain the vehicle signal, lane line signal and 4D millimeter wave radar point cloud signal, and analyze and preprocess the signal;

[0009] S2, calculating the local search radius of each point in the point cloud signal, and sorting the calculated local search radius of each point;

[0010] S3. Perform clustering processing on the point cloud signal based on the DBSCAN algorithm.

[0011] As an optimization, the vehicle signal is obtained from the vehicle CAN signal, including the vehicle's steering wheel angle, heading angular velocity, vehicle speed, and acceleration.

[0012] As an optimization, the lane line signal is obtained from the front camera CAN signal, including the cubic curve coefficient and length of the lane line.

[0013] As an optimization, the point cloud signal is obtained from the Ethernet signal of the 4D millimeter-wave radar, wherein the attributes of each point in the point cloud signal include polar diameter, horizontal angle, pitch angle, relative radial velocity, and signal-to-noise ratio.

[0014] As an optimization, the point cloud signal is preprocessed as follows:

[0015] S101, converting the coordinate information of each point in the point cloud signal into a Cartesian coordinate system;

[0016] S102, calculating the absolute radial velocity of each point;

[0017] S103: Eliminate noise points in the point cloud signal according to a preset filtering condition.

[0018] As an optimization, the absolute radial velocity is calculated by projecting the own vehicle velocity onto the angle of the corresponding point.

[0019] As an optimization, step S2 includes,

[0020] S201, select variables for calculating the distance between two points, the variables including x, y, z, and vn;

[0021] S202, determine the weight coefficient of each variable, and calculate the generalized Mahalanobis distance between two points, which is calculated as follows:

[0022] dist(p1,p2) 2 =0.8(x1-x2) 2 +1.3(y1-y2) 2 +0.1(z1-z2) 2 +(vn1-vn2) 2 (1)

[0023] Among them, x1, y1, z1 are the coordinates of point p1, x2, y2, z2 are the coordinates of point p2, and vn is the absolute radial velocity;

[0024] S203, assign a clustering factor f1 according to the absolute radial velocity of each point, and calculate it using the following formula:

[0025]

[0026] S204, assign a clustering factor f2 according to the distance between each point, and calculate it using the following formula:

[0027]

[0028] S205, set the basic local search radius ε = 1.65, calculate the final local search radius of each point, and use the following formula to calculate:

[0029] ε i =ε*f1*f2 (4)

[0030] S206. Sort the final local search radius of each point from large to small.

[0031] As an optimization, in step S3, clustering processing is performed based on breadth-first search, and the minimum number of point cloud clusters is adjusted according to a preset strategy. After the clustering processing is completed, the clustering result is output.

[0032] As an optimization, the preset strategy includes taking a larger value for a target point with a short radial distance and taking a smaller value for a target point with a long radial distance.

[0033] A storage medium stores one or more programs, and when the one or more programs are executed by a processor, the steps of the DBSCAN-based 4D millimeter-wave radar clustering method are executed.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] The present invention receives vehicle signals, lane line signals output by front camera fusion, and 4D millimeter-wave radar point cloud signals from the vehicle CAN port and the on-board Ethernet respectively, analyzes and pre-processes these signals, and then sends them to the point cloud fusion module; uses the improved DBSCAN algorithm to cluster the millimeter-wave point cloud, and fits to generate multiple observation targets in a single frame; uses the lane line information output by camera fusion to filter and delete some unimportant single-frame observation cluster targets, so that the targets output by the radar are guaranteed to be important targets that affect the automatic driving function; and performs a series of tracking algorithms such as single-sensor multi-target association, Kalman filter fusion, and target creation, deletion, and merging on the generated single-frame targets. The present invention mainly solves the problem that the traditional DBSCAN algorithm is difficult to adapt to the uneven density of 4D millimeter-wave radar point clouds for intelligent driving. The entire algorithm is based on the traditional DBSCAN algorithm, to which are added steps such as data preprocessing based on other sensors and the physical properties of each point, determining the local search radius ε for each point cloud according to the size of its physical quantity, sorting all the original detections and then clustering them, and dynamically judging whether the points in the clustering results are single-frame observation targets. By comparing the clustering results of the present invention with the traditional DBSCAN clustering method, it can be found that the clustering effect of the present algorithm is better. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is the search radius of each point of traditional DBSCAN;

[0037] Figure 2 is the local search radius of each point of DBSCAN of the present invention;

[0038] Figure 3 It is a flow chart of the clustering process of the DBSCAN algorithm of the present invention;

[0039] Figure 4This is a comparison chart of clustering effects of the DBSCAN algorithm and the traditional DBSCAN algorithm based on real driving data for 1 minute. DETAILED DESCRIPTION

[0040] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0041] Example: See Figure 1-Figure 4 .

[0042] Specifically, a 4D millimeter wave radar clustering method based on DBSCAN includes the following steps:

[0043] S1. Obtain the vehicle signal, lane line signal and 4D millimeter wave radar point cloud signal, and parse and pre-process the signal. Specifically, after connecting the hardware to the industrial computer, receive the vehicle signal from the vehicle CAN signal, mainly the steering angle, yaw rate, speed, and acceleration of the vehicle used in this algorithm; receive the lane line signal obtained by fusion from the front camera CAN signal. Since this automatic driving system is a three-lane model, there are at most six lane lines, and receive the cubic curve coefficients and their lengths of these six lane lines; receive the point cloud signal from the 4D millimeter wave radar Ethernet signal, and the maximum number of point clouds output in a single frame is 1024 points. The information contained in each point includes polar radius, horizontal angle, pitch angle, relative radial velocity, and signal-to-noise ratio SNR. These received data are parsed based on the corresponding signal protocol.

[0044] The preprocessing of point cloud signals is as follows:

[0045] S101, converting the coordinate information of each point in the point cloud signal into a Cartesian coordinate system;

[0046] S102, calculating the absolute radial velocity of each point; the absolute radial velocity is calculated by projecting the vehicle velocity on the angle of the corresponding point.

[0047] S103: Eliminate noise points in the point cloud signal according to a preset filtering condition.

[0048] Specifically, since the point cloud output by the front radar is sent in the form of polar coordinates, in order to match the subsequent algorithm, the position coordinate information of all points in the point cloud is converted from polar coordinates to Cartesian coordinates. Since the radar only sends the relative radial velocity of each point, and for better clustering effect, the clustering algorithm in this paper selects the absolute radial velocity as a weight variable for clustering, so the absolute radial velocity of each point is calculated by the projection of the vehicle speed at the corresponding point angle. In addition, since the ground is a reflection source, it is easy to cause some noise points in multipath reflection. The characteristic of these noise points is that the Z-axis position coordinate is a value less than zero and unreasonable. Therefore, the filtering conditions are set to eliminate these points. For example, in this embodiment, the installation height of the radar is 0.5m from the ground, and all detection points with a height less than -1.5m are eliminated.

[0049] S2. Calculate the local search radius of each point in the point cloud signal, and sort the calculated local search radius of each point. Specifically, the clustering flowchart is as shown in the attached figure. Figure 3 As shown in the figure, we first need to do the preparatory work before clustering, that is, to select the parameters for calculating the distance between two points and the weights between each parameter. This paper selects four parameters, x, y, z, and vn (absolute radial velocity), as variables for calculating the distance. The corresponding weights when calculating the distance are selected from these four variables. After determining the weights of each component through a large amount of data analysis, the generalized Mahalanobis distance between two points can be obtained.

[0050] S201, select variables for calculating the distance between two points, the variables including x, y, z, and vn;

[0051] S202, determine the weight coefficient of each variable, and calculate the generalized Mahalanobis distance between two points, which is calculated as follows:

[0052] dist(p1,p2) 2 =0.8(x1-x2) 2 +1.3(y1-y2) 2 +0.1(z1-z2) 2 +(vn1-vn2) 2 (1)

[0053] Among them, x1, y1, z1 are the coordinates of point p1, x2, y2, z2 are the coordinates of point p2, and vn is the absolute radial velocity;

[0054] S203, for each detection point, a unique clustering factor f1 is assigned to it with reference to its absolute radial velocity, and the judgment conditions are:

[0055]

[0056] Wherein, abs(vn) represents the absolute value function.

[0057] S204. In addition, considering that the detection density of radar is gradually sparse from near to far, which is contrary to the traditional DBSCAN algorithm, a clustering factor f2 is assigned to the distance, and its judgment conditions are:

[0058]

[0059] Among them, px means pixel.

[0060] S205, set the basic local search radius ε = 1.65, calculate the final local search radius of each point, and use the following formula to calculate:

[0061] ε i =ε*f1*f2(4)

[0062] S206. After the above operations, the local search radius of each point is determined. This step improves the problem of the traditional DBSCAN algorithm that the global variables are not easy to control the local details. However, due to the introduction of different search radii for each point, the clustering algorithm results for the same point cloud may be inconsistent. Figure 1 , 2 As shown in the figure, suppose there are two detection points A and B, the search radius of detection point A is 2, the search radius of detection point B is 1, and the generalized distance between the two points is 1.5. When searching with point A as the core point, point B is within the clustering range of point A, and when searching with point B as the core point, point A is not within the search range of point B. Assuming that point B is unable to continue searching due to the lack of an adjacent point, the clustering result will be out of control. Therefore, it is necessary to control the DBSCAN algorithm order of all points in the point cloud, sort all point clouds from large to small according to their search radius, and then perform the clustering algorithm in the sorted order to ensure that the clustering results are consistent each time. After the above processing, the point cloud data can be subjected to subsequent clustering algorithms.

[0063] S3. Perform clustering processing on the point cloud signal based on the DBSCAN algorithm. Perform traditional DBSCAN algorithm clustering processing on the point cloud. Perform clustering based on the breadth-first search method. The operation flow chart is as shown in the attached figure. Figure 3As shown in the figure, this process is the traditional DBSCAN algorithm clustering process, which will not be described in detail. After finding a point cloud cluster, the traditional DBSCAN algorithm generally defines a global minimum number of point cloud clusters, but this algorithm will adjust the minimum number of point cloud clusters to a non-fixed value based on the information of the point cloud cluster. The basic preset strategy can be that the value is larger for targets with close radial distances, and smaller for targets with long radial distances. This adjustment strategy also conforms to the detection characteristics of the front radar. After this processing, all clustering processes have been completed and the results can be output to the back end. The comparison of clustering results with real scenes and statistical indicators with traditional DBSCAN algorithms is shown in the attached figure. Figure 4 shown.

[0064] A storage medium stores one or more programs, and when the one or more programs are executed by a processor, the steps of the DBSCAN-based 4D millimeter-wave radar clustering method are executed.

[0065] The present invention receives vehicle signals, lane line signals output by front camera fusion, and 4D millimeter-wave radar point cloud signals from the vehicle CAN port and the on-board Ethernet respectively, analyzes and pre-processes these signals, and then sends them to the point cloud fusion module; uses the improved DBSCAN algorithm to cluster the millimeter-wave point cloud, and fits to generate multiple observation targets in a single frame; uses the lane line information output by camera fusion to filter and delete some unimportant single-frame observation cluster targets, so that the targets output by the radar are guaranteed to be important targets that affect the automatic driving function; and performs a series of tracking algorithms such as single-sensor multi-target association, Kalman filter fusion, and target creation, deletion, and merging on the generated single-frame targets. The present invention mainly solves the problem that the traditional DBSCAN algorithm is difficult to adapt to the uneven density of 4D millimeter-wave radar point clouds for intelligent driving. The entire algorithm is based on the traditional DBSCAN algorithm, to which are added steps such as data preprocessing based on other sensors and the physical properties of each point, determining the local search radius ε for each point cloud according to the size of its physical quantity, sorting all the original detections and then clustering them, and dynamically judging whether the points in the clustering results are single-frame observation targets. By comparing the clustering results of the present invention with the traditional DBSCAN clustering method, it can be found that the clustering effect of the present algorithm is better.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the technical solution. Those skilled in the art should understand that those modifications or equivalent substitutions of the technical solution of the present invention that do not depart from the purpose and scope of the technical solution should be included in the scope of the claims of the present invention.

Claims

1. A 4D millimeter wave radar clustering method based on DBSCAN, characterized in that: The following steps are included: S1. Obtain the vehicle signal, lane line signal and 4D millimeter wave radar point cloud signal, and analyze and preprocess the signal; the preprocessing of the point cloud signal is specifically as follows: S101, converting the coordinate information of each point in the point cloud signal into a Cartesian coordinate system; S102, calculating the absolute radial velocity of each point; the absolute radial velocity is calculated by projecting the vehicle velocity on the angle of the corresponding point; S103, removing noise points in the point cloud signal according to a preset filtering condition; S2, calculating the local search radius of each point in the point cloud signal, and sorting the calculated local search radius of each point; include, S201, select variables for calculating the distance between two points, the variables including x, y, z, and vn; S202, determine the weight coefficient of each variable, and calculate the generalized Mahalanobis distance between two points, which is calculated as follows: dist(p1,p2) 2 =0.8(x1-x2) 2 +1.3(y1-y2) 2 +0.1(z1-z2) 2 +(vn1-vn2) 2 (1) Among them, x1, y1, z1 are the coordinates of point p1, x2, y2, z2 are the coordinates of point p2, and vn is the absolute radial velocity; S203, assign a clustering factor f1 according to the absolute radial velocity of each point, and calculate it using the following formula: Where abs(vn) represents the absolute value function; S204, assign a clustering factor f2 according to the distance between each point, and calculate it using the following formula: Among them, px represents pixel; S205, set the basic local search radius ε = 1.65, calculate the final local search radius of each point, and use the following formula to calculate: e i =ε*f1*f2 (4) S206, sorting the final local search radius of each point from large to small; S3. Perform clustering processing on the point cloud signal based on the DBSCAN algorithm.

2. The 4D millimeter wave radar clustering method based on DBSCAN according to claim 1, characterized in that: The vehicle signal is obtained from the vehicle CAN signal, including the vehicle's steering wheel angle, heading angular velocity, vehicle speed, and acceleration.

3. The 4D millimeter wave radar clustering method based on DBSCAN according to claim 1, characterized in that: The lane line signal is obtained from the front camera CAN signal, including the cubic curve coefficient and length of the lane line.

4. The 4D millimeter wave radar clustering method based on DBSCAN according to claim 1, characterized in that: The point cloud signal is obtained from the Ethernet signal of the 4D millimeter wave radar, wherein the attributes of each point in the point cloud signal include polar radius, horizontal angle, pitch angle, relative radial velocity, and signal-to-noise ratio.

5. The 4D millimeter wave radar clustering method based on DBSCAN according to claim 1, characterized in that: In step S3, clustering is performed based on breadth-first search, and the minimum number of point cloud clusters is adjusted according to a preset strategy. After the clustering is completed, the clustering result is output.

6. The 4D millimeter wave radar clustering method based on DBSCAN according to claim 5, characterized in that: The preset strategy includes taking a larger value for a target point with a short radial distance and taking a smaller value for a target point with a long radial distance.

7. A storage medium, characterized in that: The storage medium stores one or more programs, and when the one or more programs are executed by the processor, the steps of the DBSCAN-based 4D millimeter-wave radar clustering method according to any one of claims 1 to 6 are executed.

Images