Millimeter wave radar multi-mode adaptive sparse point cloud clustering method based on sliding window

Abstract

The application discloses a kind of millimeter wave radar multi-mode adaptive sparse point cloud clustering methods based on sliding window, first, the lateral distance coefficient of each point is calculated to millimeter wave radar point cloud set, and DBSCAN clustering is carried out, and the clustering result is output;The effective point cloud extracted from the cluster is carried out position extrapolation set prediction, and finally sliding window processing clustering is carried out.The application can solve the clustering problem caused by the difference of point cloud under different working waveforms of millimeter wave radar, and also can solve the problem of uneven distribution of point cloud under different distances of millimeter wave radar.

Description

TECHNICAL FIELD

[0001] The application relates to a clustering method, in particular to a millimeter wave radar multi-mode adaptive sparse point cloud clustering method based on a sliding window, and belongs to the technical field of vehicle-mounted millimeter wave radars. BACKGROUND

[0002] With the development of MMIC technology in recent years, the number of millimeter wave radar channels has exploded, and after the horizontal detection capability in the early stage, the vertical detection capability has also been developed. In addition, with the increase in the number of channels, the accuracy of target detection has also been greatly improved, which enables the millimeter wave radar to also have the ability to output point clouds. The point cloud signal includes position information on the xyz axis, and also includes speed, snr, rcs and other information. In order to balance the detection resolution and the ability of long-distance detection, the 4D millimeter wave radar has to make some compromises in the waveform design. For example, Figure 1 In the specification of a 4D radar, two different wave modes are used for work.

[0003] The two modes are almost two radars in target detection. If the same clustering parameters and methods are used for clustering, the clustering effect is not good. It is easy to lose the target information of one frame. For example, Figure 2 As shown in the comparison, the point cloud is sparse in the far distance mode, and the point cloud is relatively dense in the near distance. The traditional clustering method leads to the loss of target detection in the Far mode. And when the two waveforms work alternately, the target object at the connection distance is lost in clustering.

[0004] In addition, because of the spherical propagation characteristics of electromagnetic waves, the point cloud distribution of the target detected by the millimeter wave radar is radioactive at different positions, and the clustering threshold needs to be dynamically adjusted at different distances. For example, Figure 3 As shown.

[0005] After the applicant's search, it is found that in the prior art, patent 202210460110.0 discloses a 4D millimeter wave radar clustering method based on DBSCAN and a storage medium. The method first obtains the vehicle information, lane line information and millimeter wave radar point cloud information; then calculates the local search radius of each point in the point cloud signal and sorts the search radius; finally, the point cloud is directly subjected to DBSCAN clustering processing. The invention is based on the traditional DBSCAN, adjusts and sorts the local search radius, and then clusters all the original detections and dynamically judges whether the clustering result is a single frame target. The method has simple logic, and in essence, only the search radius is adjusted according to the point cloud speed and distance. In the face of the inconsistency of the sparsity of the point cloud under different radar waveform modes, it cannot solve the inconsistency of the clustering standard. And according to the method, the parameters used in the distance judgment do not have universality, and the fixed parameters are the experience values of a specific radar. It cannot solve the millimeter wave radar clustering under different radar waveforms and different distances.

[0006] And the patent 202210562772.9 discloses a kind of millimeter wave radar point cloud target clustering method based on KDE-DBSCAN.This method is calculated according to the k value of each point cloud area for all points of single frame point cloud, corresponding associated window width W and the adjacent point RN of the point;Then according to the associated window width W and associated adjacent point RN of each store, point cloud data set D is preliminarily divided into different pseudo cluster FC;And the point number in each pseudo cluster is counted, and pseudo cluster with less than λ point number is deleted;Then the optimal associated radius and minimum associated point number of each pseudo cluster are determined;Finally, according to the optimal associated radius and minimum associated point number, DBSCAN algorithm is used to cluster each pseudo cluster respectively.The invention method can use the point cloud distribution characteristics of millimeter wave radar, and then improve the clustering quality of millimeter wave radar target.But from the basic logic of this method, the method cannot match the physical characteristics of millimeter wave radar, and conflict is prone to occur when small target objects and large target objects are close, and it cannot be distinguished.Also, occasional noise cannot be filtered out, affecting the clustering effect.At the same time, the calculation amount of multi-frame data is too large, and the clustering problem of millimeter wave radar under different waveforms cannot be solved. SUMMARY

[0007] The technical problem to be solved by the present application is to overcome the shortcomings of the prior art and provide a millimeter wave radar multi-mode adaptive sparse point cloud clustering method based on sliding window, which can solve the clustering problem caused by the difference of point cloud under different working waveforms of millimeter wave radar, and also solve the problem of uneven point cloud distribution under different distances of millimeter wave radar.

[0008] In order to achieve the above purpose, the specific technical scheme of the present application is as follows: a millimeter wave radar multi-mode adaptive sparse point cloud clustering method based on sliding window, characterized by comprising the following steps:

[0009] Step 1, coordinate conversion processing is carried out on millimeter wave radar point cloud set D, i.e., conversion to Cartesian FLU coordinate system;

[0010] Step 2, calculate the lateral distance coefficient D_rate of each point based on radar point cloud distribution formula for all points in the first millimeter wave radar point cloud set D;When DBSCAN clustering is carried out according to the D_rate, the original Euclidean distance is converted to calculate the relative equivalent distance conforming to physical characteristics;

[0011] Step 3, DBSCAN clustering is carried out on point cloud set D according to the equivalent distance in the foregoing step, and the clustering result is output;

[0012] Step 4, according to the result of DBSCAN clustering in the foregoing step, the point cloud in all clustering clusters is predicted, the next frame position of the point cloud is obtained according to the speed extrapolation, and a new point cloud set N is obtained;

[0013] Step 5: Obtain the second frame point cloud set D' according to the previous steps, and merge the new point cloud set N and point cloud set D', and perform sliding window clustering processing;

[0014] Step 6: Calculate the lateral distance coefficients for all point clouds based on the preceding steps;

[0015] Step 7: Perform a new round of DBSCAN clustering and output the results;

[0016] Furthermore, in step 2, the lateral distance coefficient d_rate is calculated as follows:

[0017]

[0018] in,

[0019]

[0020] Where g and f are adjustable coefficients, with theoretical values ​​of g = f = 1, which can be adjusted based on experience; Ri represents the distance corresponding to i, Ang_res represents the angular resolution, and r_res represents the distance resolution.

[0021] Furthermore, in step 2, the method for calculating the relative equivalent distance between two points is as follows:

[0022]

[0023] Where, x i y i Let x represent the coordinates of point i. i+1 y i+1 This represents the coordinates of point (i+1). d_rate represents the lateral distance coefficient mentioned above.

[0024] Furthermore, the specific content of step 3 is as follows:

[0025] Step 3.1: Select any data object point p in the point cloud D;

[0026] Step 3.2: For the domain parameters minimum search radius eps and minimum point cloud number minpts, calculate the dist between p and the remaining points; in the point cloud set PN, randomly select a point p' and calculate the dist between the remaining points and p'; then p and p' form a cluster;

[0027] Step 3.3: Select any one of the remaining point clouds and continue to repeat step 3.2 until all point clouds in point cloud set D have been assigned.

[0028] Further, if the point cloud set PN quantity of dist less than eps is greater than the minimum point cloud number minpts, and the RCS difference between the point clouds is less than 10dB, then p is the core point; if dist is less than eps, and the quantity is less than minpts, and rcs is less than 10dB, then p' is the core point, and so on.

[0029] Further, the specific content of step 4 is:

[0030] Step 4.1, obtain the speed and position information of all point clouds in the clustering cluster, and perform position extrapolation set prediction on the effective point cloud;

[0031] Step 4.2, extrapolate the effective point position information of the next time, and the calculation formula is as follows:

[0032] p t+1 =p t +Δt*v t (4),

[0033] And keep the remaining point cloud parameters.

[0034] Compared with the prior art, the beneficial effects of the present application are:

[0035] The present application adopts the transverse distance coefficient of the theoretical point cloud at different distances, which can obtain a clustering parameter debugging method with universality and physical characteristics of millimeter wave radar, thereby solving the natural gap of millimeter wave radar point cloud distribution at different distances, and optimizing and improving the debugging efficiency and effect of clustering. At the same time, the present application adopts the sliding window mode to deduce and predict the point cloud information of the previous frame, and combines and processes in the second frame, which solves the clustering failure caused by the great difference of point cloud under different waveforms. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 It is a comparison diagram of the far mode and the near mode of the 4D radar in the prior art.

[0037] Figure 2 It is a clustering schematic diagram of the far mode and the near mode in the prior art.

[0038] Figure 3 It is the difference of the transverse distance of the point cloud under the same angle difference at different distances.

[0039] Figure 4 It is a schematic diagram of the Cartesian FLU coordinate system in step 1 of the present application.

[0040] Figure 5 It is a flowchart of the present application.

[0041] Figure 6 It is a schematic diagram of the sliding window processing of the present application.

[0042] Figure 7 A flow chart of a traditional DBSCAN clustering process.

[0043] Figure 8 An example of clustering of the present application.

[0044] Figure 9 An explanation of extrapolated valid points of the present application. DETAILED DESCRIPTION

[0045] The technical solutions of the present application will be further described in detail below with reference to the accompanying drawings:

[0046] The flow of a millimeter wave radar multi-mode adaptive sparse point cloud clustering method based on a sliding window proposed in the embodiment is shown in Figure 5 and includes the following steps:

[0047] Step 1, process the millimeter wave radar point cloud set D, convert the position information of all point clouds into Cartesian FLU coordinate system, as shown in Figure 8 And calculate the lateral distance coefficient d_rate for all points in the first frame of millimeter wave radar point cloud set D based on the following radar point cloud distribution formula, which is only related to the distance under a certain radar:

[0048]

[0049] wherein,

[0050]

[0051] In the formula, g and f are adjustable coefficients, and the theoretical value is g = f = 1, which can be adjusted according to experience;

[0052] Ri represents the corresponding distance of i, Ang_res represents the angle resolution, and r_res represents the distance resolution.

[0053] Step 2, DBSCAN clustering of the point cloud set D, according to d_rate in the previous step, convert the original Euclidean distance when performing DBSCAN clustering, calculate the relative equivalent distance conforming to the physical characteristics, and the formula is as follows:

[0054]

[0055] Step 3, according to the adjusted equivalent distance, perform DBSCAN clustering to output the clustering result, convert the physical distance d into d_rate, as shown in Figure 8 to obtain three clusters A, B and C;

[0056] Step 3.1, randomly select a data object point p in the point cloud set D;

[0057] Step 3.2, search the minimum radius eps and the minimum point cloud number minpts of the domain parameters, calculate the dist between p and the remaining points, if the number of point cloud set PN satisfying dist < eps is greater than minpts and the RCS difference between the point clouds is less than 10 dB, then p is the core point. In the point cloud set PN, continue to select an optional point p', calculate the dist between the remaining points and p'. If the dist is less than eps and the number is greater than minpts, and the rcs is less than 10 dB, then p' is the core point, and the cycle is continued; all data object points with density reachable from p are found, and a cluster is formed;

[0058] Step 3.3, select any one of the remaining point clouds, continue the cycle of step 2.2, and all point clouds in the point cloud D are allocated, if the point cloud is allocated to a certain cluster, it is a valid point;

[0059] Step 4, extract the valid point clouds in the clustering cluster; according to the results of the DBSCAN clustering in the foregoing steps, the points in all clustering clusters are predicted, the next frame position of the points is extrapolated, and a new point cloud set N is obtained. As shown in the following formula: Figure 9 The A, B, and C clusters are extrapolated according to the positions and speeds to obtain the point cloud position information at the next time, and three point cloud clusters A', B', and C' at the new time are obtained;

[0060] Step 4.1, obtain the speed and position information of all point clouds in the clustering cluster, and perform position extrapolation set prediction on the valid point clouds;

[0061] Step 4.2, extrapolate the position information at the next time, and the calculation formula is as follows:

[0062] p t+1 = p t + Δt * v t , and the remaining point cloud parameters are retained.

[0063] Step 5, perform position extrapolation set prediction on the valid point clouds.

[0064] Step 6, after obtaining the point cloud set in the second frame, merge the extrapolated point cloud set and perform sliding window processing clustering; that is, obtain the second frame point cloud set D', and perform merging processing on the point cloud set N and the point cloud set D'.

[0065] Step 7, calculate the lateral distance coefficient of all point clouds according to the formula in step 1.

[0066] Step 8, perform a new round of DBSCAN clustering, and output the results.

[0067] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above specific embodiments, and the above specific embodiments and the description in the specification are only for further illustrating the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of the present application is defined by the claims and their equivalents.

Claims

1. A sliding window based multi-mode adaptive sparse point cloud clustering method for millimeter wave radar, characterized in that: Comprising the following steps: Step 1, coordinate conversion processing is performed on the millimeter wave radar point cloud set D, that is, conversion to a Cartesian FLU coordinate system; Step 2, all points in the first frame millimeter wave radar point cloud set D are calculated based on the radar point cloud distribution formula, and the lateral distance coefficient D_rate of each point is calculated; when performing DBSCAN clustering according to the D_rate, the original Euclidean distance is converted, and the relative equivalent distance meeting the physical characteristics is calculated; the calculation method of the lateral distance coefficient d_rate is as follows: （1） Wherein, （2） Wherein, g and f are adjustable coefficients, and the theoretical value is g=f=1, which can be adjusted according to experience; Ri represents the corresponding distance of i, Ang_res represents the angle resolution, and r_res represents the distance resolution; The calculation method of the relative equivalent distance is as follows: （3） wherein x i , y i represent the coordinates of point i, x i+1 , y i+1 represent the coordinates of point (i+1); d_rate represents the aforementioned transverse distance coefficient; Step 3, DBSCAN clustering is performed on the point cloud set D according to the equivalent distance in the foregoing step, and the clustering result is output, and the specific steps are as follows: Step 3.1, an arbitrary data object point p is selected in the point cloud set D; Step 3.2, the domain parameters minimum search radius eps and minimum point cloud number minpts are calculated, and the dist between p and the remaining points is calculated; an arbitrary point p' is selected in the point cloud set PN, and the dist between the remaining points and p' is calculated; a cluster is formed by p and p'; Step 3.3, an arbitrary point in the remaining point cloud is selected, and step 3.2 is continued to cycle until all point clouds in the point cloud set D are assigned; Step 4, according to the result of DBSCAN clustering in the foregoing step, the point clouds in all clusters are predicted, the next frame position of the point cloud is obtained according to the speed extrapolation, and a new point cloud set N is obtained; Step 5, the second frame point cloud set D' is obtained according to the foregoing step, and the new point cloud set N and the point cloud set D' are merged, and a sliding window clustering process is performed; Step 6, the lateral distance coefficient of all point clouds is calculated according to the foregoing step; Step 7, a new round of DBSCAN is performed, and the result is output.

2. The sliding window based millimeter wave radar multi-mode adaptive sparse point cloud clustering method according to claim 1, characterized in that: If the number of point cloud set PN whose dist is less than eps is greater than the minimum point cloud number minpts, and the RCS difference between the point clouds is less than 10 dB, then p is a core point; If the dist is less than eps, and the number is less than minpts, and the rcs is less than 10 dB, then p' is a core point, and the cycle is continued.

3. The sliding window based mmWave radar multi-mode adaptive sparse point cloud clustering method of claim 1, wherein: The specific content of step 4 is as follows: Step 4.1, the speed and position information of all point clouds in the clustering cluster are obtained, and the position extrapolation set prediction is performed on the effective point cloud; Step 4.2, the effective point position information at the next time is extrapolated, and the calculation formula is as follows: (4)， And the remaining point cloud parameters are retained.

Images