Uneven density clustering method of vehicle-mounted lidar point cloud based on OPTICS algorithm

Abstract

This invention discloses a method for clustering uneven density of vehicle-mounted lidar point clouds based on the OPTICS algorithm. The method comprises the following steps: 1) generating a decision sequence diagram; 2) calculating the rate of change of the decision sequence diagram; 3) screening the peak points of the rate of change of the decision sequence diagram; 4) reorganizing and sorting the decision sequence diagram; and 5) using a threshold function to divide and cluster. This method achieves adaptive density clustering, avoiding the drawbacks of poor clustering of sparse point clouds at the periphery when the density threshold is set high, and cluster merging of dense point clouds at the interior when the threshold is set low.

Description

Technical Field

[0001] The present invention relates to a laser radar point cloud clustering technology, specifically a vehicle-mounted laser radar point cloud uneven density clustering method based on the OPTICS (Ordering points to identify the clustering structure, abbreviated as OPTICS) algorithm. Background Art

[0002] LiDAR technology uses lasers as signal carriers for detection, quickly and accurately acquiring spatial information about targets and achieving synchronization of three-dimensional coordinates. In the field of autonomous driving, automotive LiDAR is primarily used for perception, measuring the distance to surrounding objects and accurately sensing the surrounding driving environment to ensure the full performance of autonomous driving algorithms.

[0003] Clustering is the process of dividing a collection of objects into multiple clusters with shared characteristics, with objects in the same cluster being similar and different clusters being different. LiDAR point cloud clustering can be used to create clearer spatial information. In the field of autonomous driving, clustering helps the sensor integrate with other sensors to distinguish objects, vehicles, people, and other information within the point cloud. Numerous methods have been developed for cluster analysis, including partition-based, hierarchical, and density-based methods. Density-based methods are more suitable for processing LiDAR point cloud information, overcoming the drawback of partition-based methods, which only reveal "circular" clusters.

[0004] Among the representative density-based clustering algorithms, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise, DBSCAN for short) algorithm uses a given radius and a minimum number of objects as parameters to divide areas with sufficiently high density into clusters. The OPTICS algorithm is an improved version of DBSCAN. Compared with DBSCAN, it reduces the sensitivity of input parameters. The OPTICS algorithm sorts the objects in the data set to obtain an ordered list of objects. Through this ordered list, a decision diagram can be obtained. The decision diagram can be used to understand the clustering situation when the radius takes a specific value. However, the above two methods are not ideal for clustering point clouds with sparse distribution on the periphery and dense distribution inside the vehicle-mounted radar. Summary of the Invention

[0005] The present invention addresses the shortcomings of existing technologies by providing a method for clustering uneven density point clouds from vehicle-mounted LiDAR radars based on the OPTICS algorithm. This method enables adaptive density clustering, avoiding the drawbacks of poor clustering of sparse point clouds at the periphery when the density threshold is set high, and cluster merging of dense point clouds at the interior when the threshold is set low.

[0006] The technical solution for achieving the purpose of the present invention is:

[0007] A vehicle-mounted lidar point cloud uneven density clustering method based on the OPTICS algorithm includes the following steps:

[0008] 1) Generate decision sequence diagram: Perform OPTICS clustering on the preprocessed point cloud to generate a decision sequence diagram. Preprocessing uses the RANSAC plane segmentation method to remove the ground and voxel downsampling. Preprocessing can reduce the complexity of the point cloud. The horizontal axis of the decision sequence diagram is the processing order and the vertical axis is the reachable distance of the point. Assuming that the point cloud consists of N points, the decision diagram is N points and their corresponding reachable distances R = {r1, r2, ... r N}} is the discrete function of the mapping;

[0009] 2) Calculate the change rate of the decision sequence diagram: Generate a decision diagram G = {g1, g2, ... g N As shown in formula (1):

[0010]

[0011] Where j is the step size of discrete differentiation, and G can be approximately regarded as the rate of change of R;

[0012] 3) Filter the peak point of the change rate of the decision sequence diagram: find the position p of the local peak point that meets the conditions for G. Assuming that there are M peak points that meet the conditions, we get G peaks ={p1,p2...p M As the area division point set, h is the required peak height. The value here is related to the change amplitude of the decision sequence diagram. The reference value is 1 / 3 of the amplitude peak. Peak points with insufficient height are screened. dis is the minimum horizontal distance required in the sample. The value here is related to the size of the point cloud data. The reference value here is N / 20. The highest peak value is taken within the dis range of each point until all peak points meet the conditions. Different data have different values. The appropriate value can avoid the division points being too dense and affecting the clustering effect, as shown in formula (2):

[0013] p j =g i (g i >h,g i =max(g i-dis ...g i ...g i+dis )) (2),

[0014] where g i-dis =g1(i<1+dis), g i+dis =g N (i>N-dis)

[0015] 4) Reorganization and sorting of decision sequence diagram: according to the peak point set G peaks The M points divide R into M+1 regions. Assume that the peak point p j The corresponding reachable distance point is r oj , calculate the mean of each area, and then sort the areas from small to large according to the mean, and reorganize them into a new point set R sort , as shown in formula (3):

[0016]

[0017] 5) Threshold function clustering: The R obtained in step 4) sort Compared with the decision graph of R processed in step 1), the decision graph has the distribution characteristics of high density in the front and low density in the back. Therefore, compared with the traditional optics clustering, a fixed threshold ε is selected for R to extract clusters, and a threshold function ε(i) with low values ​​in the front and high values ​​in the back is used to extract clusters for R. sort Processing, such as linear function, exponential function, as shown in formula (4): Then adaptive density clustering can be achieved. Here, the threshold is the reachable distance threshold, and the relationship with the density threshold is:

[0018]

[0019] This technical solution reorganizes the decision graph without affecting the OPTICS clustering effect, forming a new decision graph in which the dense point cloud part is sorted first and the sparse point cloud part is sorted last, and uses an adaptive threshold function for clustering grouping.

[0020] This method can achieve adaptive density clustering, that is, it can avoid the defects of poor clustering effect of sparse point clouds in the periphery when the density threshold is set high and cluster merging of dense point clouds in the interior when the threshold is set low. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of the comparison between the original point cloud and the pre-processed point cloud in the embodiment;

[0022] Figure 2 The decision graph, change rate graph, and reorganized decision graph generated by density clustering OPTICS in the embodiment;

[0023] Figure 3 Schematic diagram of the process of the embodiment method;

[0024] Figure 4 Schematic diagram comparing clustering results under the conditions of a small ε threshold, a large ε threshold, and ε threshold function processing in the embodiment;

[0025] Figure 5 Schematic diagram of comparison of the number of clusters under the processing of a small ε threshold, a large ε threshold, and an ε threshold function in the embodiment. DETAILED DESCRIPTION

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the present invention is not limited thereto.

[0027] Example:

[0028] Reference Figure 1 、 Figure 3 , a vehicle-mounted lidar point cloud uneven density clustering method based on the OPTICS algorithm, including the following steps:

[0029] 1) Generate decision sequence diagram: Perform OPTICS clustering on the pre-processed point cloud to generate a decision sequence diagram. The decision sequence diagram is as follows: Figure 2 As shown in the figure above, the preprocessing method is to remove the ground and voxel downsampling by using the RANSAC plane segmentation method. Preprocessing can reduce the complexity of the point cloud. The horizontal axis in the decision sequence diagram is the processing order and the vertical axis is the reachable distance of the point. Assuming that the point cloud consists of N points, the decision diagram is N points and their corresponding reachable distances R = {r1, r2, ... r N}} is the discrete function of the mapping;

[0030] 2) Calculate the change rate of the decision sequence diagram: Generate a decision diagram G = {g1, g2, ... g N As shown in formula (1):

[0031]

[0032] Where j is the step size of discrete differentiation, G can be approximately regarded as the rate of change of R, such as Figure 2 As shown in the middle picture;

[0033] 3) Filter the peak point of the change rate of the decision sequence diagram: find the position p of the local peak point that meets the conditions for G. Assuming that there are M peak points that meet the conditions, we get G peaks ={p1,p2...p M} is used as the region partition point set, h is the required peak height, the value in this example is related to the change amplitude of the decision sequence diagram, and the reference value is 1 / 3 of the peak amplitude. Filter out peak points with insufficient height, dis is the minimum horizontal distance required in the sample, the value in this example is related to the size of the point cloud data, and the reference value here is N / 20. The highest peak value within the dis range of each point is taken until all peak points meet the conditions. Different data have different values. The appropriate value can avoid the partition points being too dense, which affects the clustering effect, as shown in formula (2):

[0034]

[0035] where g i-dis =g1(i<1+dis), g i+dis =g N (i>N-dis)

[0036] 4) Reorganization and sorting of decision sequence diagram: according to the peak point set G peaks The M points divide R into M+1 regions. Assume that the peak point p j The corresponding reachable distance point is r oj , calculate the mean of each area, and then sort the areas from small to large according to the mean, and reorganize them into a new point set R sort , as shown in formula (3):

[0037]

[0038] 5) Threshold function clustering: The R obtained in step 4) sort The decision diagram of step 1) is the same as the R decision diagram processed, such as Figure 2 As shown, the distribution characteristics are high in the front and low in the back. Compared with the traditional optics clustering, the fixed threshold ε is selected for R to extract clusters. The threshold function ε(i) with low values ​​in the front and high values ​​in the back is used to extract clusters for R. sort Processing, for example, is shown in formula (4): Then adaptive density clustering can be achieved. Here, the threshold is the reachable distance threshold, and the relationship with the density threshold is:

[0039]

[0040] Verification using vehicle-mounted lidar point clouds from the nuscenes dataset demonstrates that, compared to traditional approaches, this method can avoid the drawbacks of poor clustering of sparse point clouds at the periphery when the distance density threshold is set high, and cluster merging of dense point clouds at the interior when the threshold is set low.

[0041] like Figure 4 As shown in the figure, the clustering result marks the sparse points above the distance threshold as discrete points in black. It can be seen that when the ε threshold is small, the sparse outer part is identified as a discrete point, while when the ε threshold is large, the sparse outer part is clustered, but the classes of the dense inner part that are close to each other are merged into one class. In the case of the ε threshold function, reasonable clustering is obtained both inside and outside.

[0042] like Figure 5As shown in the figure, the number of clusters in the three cases is compared. It can be seen that when the ε threshold is small, the sparse classes are divided into multiple classes, so the number of clusters is large. When the ε threshold is large, the close classes are merged, so the number of clusters is small. The appropriate ε threshold function can obtain reasonable clustering, which indirectly reflects the rationality of the clustering effect of this method.

Claims

1. A vehicle-mounted lidar point cloud uneven density clustering method based on the OPTICS algorithm, characterized by: The steps include: 1) Generate decision sequence diagram: Perform OPTICS clustering on the preprocessed point cloud to generate a decision sequence diagram. The preprocessing method is to remove the ground and voxel downsampling using the RANSAC plane segmentation method. The horizontal axis of the decision sequence diagram is the processing order and the vertical axis is the reachable distance of the point. Assuming that the point cloud consists of N points, the decision diagram is N points and their corresponding reachable distances R = {r1, r2, ... r N }} is the discrete function of the mapping; 2) Calculate the change rate of the decision sequence diagram: Generate a decision diagram G = {g1, g2, ... g N As shown in formula (1): Where j is the step size of discrete differentiation, and G is approximately regarded as the rate of change of R; 3) Filter the peak point of the change rate of the decision sequence diagram: find the position p of the local peak point that meets the conditions for G. Assuming that there are M peak points that meet the conditions, we get G peaks ={p1,p2...p M } is used as the area division point set, h is the required peak height, which is 1 / 3 of the amplitude peak value, and the peak points with insufficient height are screened out. dis is the minimum horizontal distance required in the sample. The value here is related to the size of the point cloud data and is N / 20. The highest peak value within the dis range of each point is taken until all peak points meet the conditions, as shown in formula (2): p j =g i (g i >h,g i =max(g i-dis ...g i ...g i+dis )) (2), where g i-dis =g1(i<1+dis), g i+dis =g N (i>N-dis) 4) Reorganization and sorting of decision sequence diagram: according to the peak point set G peaks The M points divide R into M+1 regions. Assume that the peak point p j The corresponding reachable distance point is Calculate the mean of each area, then sort the areas from small to large according to the mean, and reorganize them into a new point set R sort , as shown in formula (3): 5) Threshold function clustering: The R obtained in step 4) sort Compared with the decision graph of R processed in step 1), the decision graph has the distribution characteristics of high density in the front and low density in the back. Compared with the traditional optics clustering, the fixed threshold ε is selected for R to extract clusters. The threshold function ε(i) with low value in the front and high value in the back is used to extract R. sort Processing is performed as shown in formula (4): Adaptive density clustering is achieved, where the threshold is the reachable distance threshold, and the relationship with the density threshold is:

Images