Point cloud data segmentation method and system based on DBSCAN clustering algorithm

Abstract

The invention provides a point cloud data segmentation method and system based on a DBSCAN clustering algorithm, and belongs to the technical field of data processing, and the method comprises the steps: obtaining non-ground point cloud data, and obtaining final high-density data and final low-density data; obtaining a DBSCAN clustering algorithm under a plurality of parameter combinations, obtaining a first reference value and a second reference value within the fitness of a clustering result under each parameter combination according to the clustering cluster, the core point and the noise point under each parameter combination, and further obtaining an initial clustering cluster and an initial noise point; and clustering the initial noise points to obtain a final clustering result. According to the invention, a secondary clustering and distance verification mode is adopted, so that a low-density object is prevented from being misjudged as noise, and missed classification is reduced; according to the non-ground point cloud segmentation method, through a processing mode of first partitioning, then targeted clustering and finally verification optimization, the influence of non-uniform point cloud density on DBSCAN clustering is effectively overcome, and the accuracy of non-ground point cloud segmentation is improved.

Description

Technical Field

[0001] The present invention belongs to the field of data processing, and in particular relates to a point cloud data segmentation method and system based on a DBSCAN clustering algorithm. Background Art

[0002] Point cloud data is a collection of discrete three-dimensional points acquired through 3D scanning equipment. Each point typically contains coordinate information and may also include attributes such as color, intensity, and normal vectors. It is widely used in fields such as computer vision, robotics, and autonomous driving. Because road surface point cloud data contains complex mixed scene information, it is necessary to segment the road surface point cloud data. Currently, when segmenting point cloud data using the RANSAC algorithm, it only divides the point cloud data into ground point cloud data and non-ground point cloud data, and cannot further classify the non-ground point cloud data. Therefore, clustering is performed on the non-ground point cloud data.

[0003] Cluster analysis is a fundamental technique in data mining. It aims to partition objects in a dataset into multiple groups based on similarity, achieving high similarity within groups and significant differences between them. The density-based clustering algorithm, DBSCAN, dynamically determines the number of clusters based on changes in the density of the dataset, eliminating the need to pre-define the number of clusters. It can also identify clusters of any shape and size, making it suitable for further classification of non-ground point cloud data. When acquiring road point cloud data, the closer an object is to the 3D point cloud data acquisition device, the more 3D point cloud data for that object is collected, resulting in point clouds containing data of varying densities. This makes it impossible to achieve optimal clustering results using a single combination of neighborhood radius and minimum number of parameters when clustering point cloud data using the DBSCAN algorithm. Furthermore, the neighborhood radius and minimum number of points must be manually configured when clustering data using the DBSCAN algorithm. This results in segmentation results that differ significantly from actual results when using the DBSCAN algorithm for non-ground point cloud data. Summary of the Invention

[0004] In order to solve the problem that when non-ground point cloud data is currently classified using the DBSCAN clustering algorithm, the clustering results are too different from the actual non-ground point cloud object classification, the present invention provides a point cloud data segmentation method based on the DBSCAN clustering algorithm.

[0005] In order to achieve the above object, the present invention provides the following technical solutions: Acquire non-ground point cloud data; According to the distribution of non-ground point cloud data, the non-ground point cloud data is divided into final high-density data and final low-density data; While ensuring that the core points are the final high-density data, the DBSCAN clustering algorithm is used to cluster all non-ground point cloud data to obtain the initial clusters and initial noise points; While ensuring that the core points are the final low-density data, the DBSCAN clustering algorithm is used to cluster the initial noise points to obtain noise clusters and secondary noise points; According to the Euclidean distance between the secondary noise points and the initial clusters and each data in the noise clusters, the final clustering result is obtained, and the segmentation of the non-ground point cloud data is completed.

[0006] Based on point cloud distribution characteristics, this invention classifies non-ground point clouds into "final high-density data" (dense points at close range, such as nearby objects) and "final low-density data" (sparse points at far range, such as distant objects) according to a density threshold. This overcomes the limitation of "single parameter processing of full-density data," provides a basis for subsequent parameterized clustering, and is a prerequisite for addressing density unevenness. The "final high-density data" is clustered using parameters appropriate for dense points to obtain initial clusters (primarily corresponding to nearby objects) while simultaneously marking initial noise points (which may include low-density points or true noise). High-density areas are focused on capturing clear nearby objects first, avoiding excessive merging of dense points due to overly broad parameters. The "initial noise points" from the initial clustering are filtered out to obtain "final low-density data," and secondary clustering is performed using parameters appropriate for sparse points to obtain noise clusters (corresponding to distant, sparse objects). For low-density areas missed by the initial clustering, parameter adaptation is used to identify distant objects, avoiding missed classifications due to overly strict parameters. Furthermore, the Euclidean distance between the secondary noise points (points not secondary clustered) and the initial cluster and the noise cluster is calculated. Distance verification is used to compensate for the boundary error of the two clusters and ensure the correct classification of edge points and sparse points.

[0007] Preferably, the non-ground point cloud data is divided into final high-density data and final low-density data according to the distribution of the non-ground point cloud data, and the specific steps are as follows: Calculate the The Euclidean distance between the first non-ground point cloud data and all non-ground point cloud data is obtained. A distance set of non-ground point cloud data; The first All data corresponding to the minimum distance in the distance set of non-ground point cloud data are recorded as The nearest neighbor data of non-ground point cloud data; Through the Non-ground point cloud data and its nearest neighbor data, and get a nearest neighbor data pair; The Euclidean distance between the two data in each nearest neighbor data pair is recorded as the nearest neighbor distance of the nearest neighbor data pair, and the nearest neighbor distance set is obtained; Each neighbor distance in the neighbor distance set is used as a segmentation threshold, and high-density data and low-density data under each segmentation threshold are obtained according to the segmentation threshold; According to the high-density data and low-density data in each neighbor data pair under each segmentation threshold, the final high-density data and the final low-density data are obtained.

[0008] Preferably, the specific steps of obtaining high-density data and low-density data under each segmentation threshold according to the segmentation threshold are as follows: Let the segmentation threshold be The non-ground point cloud data pairs within the nearest neighbor data pairs whose nearest neighbor distance is greater than the segmentation threshold are recorded as low-density data pairs; The non-ground point cloud data pairs within the nearest neighbor data pairs whose neighbor distance is less than or equal to the segmentation threshold are recorded as high-density data pairs.

[0009] Preferably, the specific steps of obtaining the final high-density data and the final low-density data according to the high-density data and the low-density data in each neighboring data pair under each segmentation threshold are as follows: Get the segmentation threshold as The segmentation effectiveness parameter when the neighbor distance is 0 is calculated as follows: ; Where, Indicates that the segmentation threshold is The segmentation effectiveness parameter when the neighbor distance is , represents the number of data points belonging to high-density data pairs, represents the number of data points belonging to low-density data pairs, For the The number of nearest neighbor data point pairs whose nearest neighbor distance is less than the segmentation threshold in all nearest neighbor data pairs of the data points belonging to the high-density data pairs; For the The number of nearest neighbor data point pairs whose nearest neighbor distance is greater than the segmentation threshold in all nearest neighbor data pairs of the data points belonging to the low-density data pairs; The nearest neighbor distance corresponding to the maximum segmentation effectiveness is recorded as the optimal segmentation threshold; The non-ground point cloud data within the nearest neighbor data pair whose neighbor distance is greater than the optimal segmentation threshold is recorded as the final low-density data; the non-ground point cloud data within the nearest neighbor data pair whose neighbor distance is less than or equal to the optimal segmentation threshold is recorded as the final high-density data.

[0010] Preferably, it also includes combining the non-dominated sorting genetic algorithm with the DBSCAN clustering algorithm to obtain the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm, and the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm to obtain the final segmentation result.

[0011] Preferably, the optimal parameter combination when clustering the non-ground point cloud data using the DBSCAN clustering algorithm obtains the final segmentation result, and the specific steps are as follows: A non-dominated sorting genetic algorithm is used to obtain multiple parameter combinations; each parameter combination includes a high-density data parameter combination of neighborhood radius and minimum number of points, and a low-density parameter combination of neighborhood radius and minimum number of points, and the neighborhood radius and minimum number of points in the high-density parameter combination are different from those in the low-density parameter combination; The neighborhood radius and minimum number of points under the high-density data parameter combination in the y-th parameter combination are used as the neighborhood radius and minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the high-density parameter combination in the y-th parameter combination is obtained; The neighborhood radius and minimum number of points under the low-density data parameter combination in the y-th parameter combination are used as the neighborhood radius and minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the low-density parameter combination in the y-th parameter combination is obtained; Under the condition that the core point is the final high-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the high-density parameter combination within the y-th parameter combination is used to cluster all non-ground point cloud data, and the cluster clusters, core points and noise points under the high-density parameter combination within the y-th parameter combination are obtained, which are recorded as the high-density cluster cluster within the y-th parameter combination, the noise points and core points within the parameter combination; Under the condition that the core points are the final low-density data, the DBSCAN clustering algorithm under the low-density parameter combination in the y-th parameter combination is used to continue clustering the noise points in the y-th parameter combination. The clusters, noise points and core points obtained by this clustering are recorded as the low-density clusters in the y-th parameter combination, the secondary noise points in the y-th parameter combination and the core points in the y-th parameter combination respectively. Combine the yth parameter into the The Euclidean distance between the secondary noise point and each data in the initial cluster and the noise cluster is recorded as the yth parameter combination. The difference distance of the secondary noise points; If the yth parameter combination If the minimum difference distance of the secondary noise points is less than or equal to the neighborhood radius of the high-density data parameter combination within the yth parameter combination, then the yth parameter combination within the yth parameter combination The secondary noise points are placed in the cluster corresponding to the data with the minimum difference distance; The above operation is performed on all secondary noise points in the yth parameter combination to obtain the updated low-density clusters and high-density clusters in the yth parameter combination, and they are recorded as the final clusters in the yth parameter combination; Obtaining a first reference value of the final clustering result in the yth parameter combination through the distribution of the core point in the final cluster and the data around the core point in the yth parameter combination; According to the distances of the core points in different final clustering clusters in the yth parameter combination, the distribution differences of the data around the core points, and the distribution of the noise points, a second reference value of the final clustering result in the yth parameter combination is obtained; The first reference value and the second reference value of the final clustering result in each parameter combination are input into the non-dominated sorting genetic algorithm as the two fitness values ​​of the parameter combination to obtain the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm; The neighborhood radius and the minimum number of points in the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm are used as the neighborhood radius and the minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the high-density parameter combination in the optimal parameter combination is obtained; While ensuring that the core points are the final high-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the high-density parameter combination within the optimal parameter combination is used to cluster all non-ground point cloud data. The cluster clusters and noise points under the high-density parameter combination within the optimal parameter combination are obtained, which are recorded as the initial cluster clusters and initial noise points respectively. The neighborhood radius and minimum number of points under the low-density data parameter combination within the optimal parameter combination when clustering all non-ground point cloud data are set as the neighborhood radius and minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the low-density parameter combination within the optimal parameter combination is obtained; While ensuring that the core points are the final low-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the low-density parameter combination within the optimal parameter combination is used to cluster all non-ground point cloud data. The cluster clusters and noise points under the low-density parameter combination within the optimal parameter combination are obtained, which are recorded as noise cluster clusters and secondary noise points respectively. The final clustering result is obtained based on the Euclidean distance between the secondary noise point and the initial cluster and each data in the noise cluster.

[0012] Preferably, the first reference value of the final clustering result in the yth parameter combination is obtained by the distribution of the core point in the final cluster and the data around the core point in the yth parameter combination, and the specific steps are as follows: Combined with the yth parameter The final cluster within the The data whose Euclidean distance of the core points is less than or equal to the neighborhood radius under the high-density parameter combination in the y-th parameter combination is recorded as the y-th parameter combination. The final cluster within the Neighborhood points of the core point; Combine the yth parameter into the The final cluster within the The mean of the Euclidean distances between the core point and all its neighboring points is used as the The final cluster within the The density estimate of the core points; Combine the yth parameter into the The variance of the density estimation value of all core points in the final cluster is recorded as the variance of the density estimation value of the yth parameter combination. The density consistency of the final clusters; The sum of the density consistencies of all final clustering clusters in the yth parameter combination is recorded as the first reference value of the final clustering result in the yth parameter combination.

[0013] Preferably, the second reference value of the final clustering result in the yth parameter combination is obtained according to the distances between the core points in different final clustering clusters in the yth parameter combination, the distribution differences of the data around the core points, and the distribution of the noise points. The specific steps are as follows: Calculate the yth parameter combination Each core point in the final cluster is The distance between each core point in the final cluster; If the yth parameter combination The final cluster within the The first core point in the yth parameter combination The final cluster within the If the Euclidean distance between the core points is less than 2 times the neighborhood radius under the high-density parameter combination in the y-th parameter combination, the two clusters are considered adjacent. Then the difference between the Euclidean distance between the core points in the two different clusters and 2 times the neighborhood radius under the high-density parameter combination in the y-th parameter combination is used as a parameter for the over-segmentation score; Get the yth parameter combination The final clusters and The specific calculation formula for the over-segmentation score of the final cluster is as follows: ; Where, The yth parameter combination The final clusters and The over-segmentation score of the final clusters, Indicates the yth parameter combination The variance of the density estimate of the core points in the final cluster, Indicates the yth parameter combination The variance of the density estimates of the cores within the final clusters, Indicates the yth parameter combination The final clusters and The number of over-subdivisions of the final clusters, Represents the neighborhood radius under the high-density data parameter combination within the yth parameter combination, Indicates the yth parameter combination The final clusters and The final cluster The degree of over-detailed performance, represents the absolute value function; The sum of the over-segmentation scores of all two different final clusters in the yth parameter combination is input into the normalization function, and the output is recorded as the over-segmentation score of the final clustering result in the yth parameter combination; The calculation formula for obtaining the noise ratio of the final clustering result in the yth parameter combination is as follows: ; Where, Indicates the noise ratio of the final clustering result in the y-th parameter combination, Indicates the number of non-ground point cloud data, Indicates the number of non-ground point cloud data that are not in the final cluster within the y-th parameter combination; The specific calculation formula for obtaining the second reference value within the fitness of the clustering result under the high-density parameter combination within the y-th parameter combination is as follows: ; Where, The second parameter representing the final clustering result in the y-th parameter combination, represents the over-segmentation score of the final clustering result in the y-th parameter combination, Indicates the noise proportion of the final clustering result within the y-th parameter combination.

[0014] Preferably, the final clustering result is obtained according to the Euclidean distance between the secondary noise point and the initial cluster and each data in the noise cluster, and the specific steps are as follows: The first The Euclidean distance between the secondary noise point and each data in the initial cluster and the noise cluster is recorded as The difference distance of the secondary noise points; Jordi If the minimum difference distance of the secondary noise points is less than or equal to the neighborhood radius of the DBSCAN clustering algorithm under the optimal high-density parameter combination when clustering all non-ground point cloud data, then the The secondary noise points are placed in the cluster corresponding to the data with the minimum difference distance; The above operation is performed on all secondary noise points to obtain updated initial clusters and updated noise clusters; the final clustering results are obtained based on the updated initial clusters and noise clusters.

[0015] The present invention also proposes a point cloud data segmentation system based on the DBSCAN clustering algorithm, comprising: Point cloud data acquisition module, used to obtain non-ground point cloud data; A point cloud data segmentation module is used to divide the non-ground point cloud data into final high-density data and final low-density data according to the distribution of the non-ground point cloud data; While ensuring that the core points are the final high-density data, the DBSCAN clustering algorithm is used to cluster all non-ground point cloud data to obtain the initial clusters and initial noise points; While ensuring that the core points are the final low-density data, the DBSCAN clustering algorithm is used to cluster the initial noise points to obtain noise clusters and secondary noise points; According to the Euclidean distance between the secondary noise points and the initial clusters and each data in the noise clusters, the final clustering result is obtained, and the segmentation of the non-ground point cloud data is completed.

[0016] The point cloud data segmentation method based on the DBSCAN clustering algorithm provided by the present invention has the following beneficial effects: when clustering non-ground point cloud data, the present invention first clusters the non-ground point cloud data when obtaining the final clustering result, based on the feature that when the point cloud data of the object is obtained by the laser radar, the amount of point cloud data of the object is inversely proportional to the distance of the laser radar. This solves the problem that the amount of point cloud data of the object is inversely proportional to the distance of the laser radar, that is, the closer the distance of an object to the laser radar, the more laser point cloud images of the object, so that when clustering is performed directly through the DBSCAN clustering algorithm, some point cloud data of objects that are far away from the laser radar will be identified as noise; based on the feature that when clustering data points through the DBSCAN clustering algorithm, the data points are essentially clustered according to the distribution of other data points around each data point, the distribution of data points around the core point in each cluster cluster is used as a parameter for evaluating the effectiveness of the clustering results, which solves the current problem of evaluating clustering results. The effectiveness method is not suitable for evaluating the clustering results obtained by the DBSCAN algorithm; evaluating the quality of clustering by the distribution of data points around the core point in each cluster may cause over-clustering. The second parameter for evaluating the effectiveness of the clustering results is obtained by the distance distribution of core points in different clusters, which avoids the occurrence of over-clustering; then, the two parameters of the effectiveness of the clustering results when clustering non-ground point cloud data under each parameter combination are used to obtain the best clustering result in the initial clustering using the non-dominated sorting genetic algorithm, and then the noise points in the initial clustering result are clustered again. This solves the problem that when point cloud data is obtained by lidar, the amount of point cloud data of the object is inversely proportional to the distance between the lidars, which makes the clustering results under the current method differ greatly from the actual situation, and the existing clustering result evaluation criteria cannot well evaluate the clustering results when clustering data through the DBSCAN clustering algorithm, resulting in the current segmentation effect of point cloud data not being ideal when segmenting using the DBSCAN clustering method. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] To more clearly illustrate the embodiments of the present invention and its design, the following briefly introduces the drawings required for this embodiment. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort.

[0018] Figure 1 Flowchart of a point cloud data segmentation method based on the DBSCAN clustering algorithm according to Example 1 of the present invention; Figure 2 is the real scene image of the target area; Figure 3is the point cloud image of the target area; Figure 4 is the initial segmentation image of the point cloud image of the target area; Figure 5 This is a schematic diagram of the results when clustering data using the DBSCAN clustering algorithm; Figure 5 (a) is the clustering result when the data is clustered using the DBSCAN clustering algorithm. Figure 5 (b) is the clustering result when the data is clustered using the traditional clustering algorithm; Figure 6 Schematic diagram of the over-clustering result; Figure 7 is a schematic diagram of the distribution of noise points and data points within the cluster after clustering; Figure 7 (a) is the clustering result image of the preliminary clustering of noise points. Figure 7 (b) is for Figure 7 The clustering result image after the noise points in (a) are redistributed; Figure 8 This is the result of clustering non-ground point cloud data using existing methods; Figure 9 This is a result diagram of clustering non-ground point cloud data using the method of the present invention; Figure 10 This is the result diagram of the division of high and low density data. DETAILED DESCRIPTION

[0019] In order to enable those skilled in the art to better understand the technical solution of the present invention and to be able to implement it, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0020] Example 1 The present invention provides a point cloud data segmentation method based on the DBSCAN clustering algorithm, specifically Figure 1 As shown, including: Step S001: Acquire non-ground point cloud data.

[0021] Specifically, a laser radar is used to obtain a laser point cloud image of a target area on the road surface, and the RANSAC algorithm is used to divide the laser point cloud image of the target area into non-ground point cloud data and ground point cloud data. Figure 2 is the real scene image of the target area, Figure 3 is the point cloud image of the target area, Figure 4 is the initial segmentation image of the point cloud image of the target area. And Figure 2 The objects contained in the left box in Figure 3The objects contained in the left box are all the same car. Figure 2 The objects contained in the second left box are Figure 3 The objects contained in the second left box are the same person. Figure 4 The red pixels in the figure represent non-ground point cloud data, the blue pixels represent ground point cloud data, and the white areas represent blind spots of the LiDAR, where the LiDAR does not obtain point cloud data. The use of the RANSAC algorithm to separate point cloud data into ground and non-ground point cloud data is a well-known technique and will not be described in detail in this embodiment.

[0022] Step S002: Based on the distribution of the non-ground point cloud data, several neighboring data pairs and segmentation thresholds are obtained, and then high-density data and low-density data under each segmentation threshold are obtained; based on the high-density data and low-density data within each neighboring data pair under each segmentation threshold, the non-ground point cloud data is divided into final high-density data and final low-density data.

[0023] It's important to note that when acquiring 3D point cloud data using LiDAR, the density of an object's 3D point cloud data is related to the distance between the object and the LiDAR. Specifically, when acquiring 3D point cloud data using LiDAR, the distance between the collected 3D point cloud data and the LiDAR is characterized by being denser near and sparser far away. Therefore, when clustering non-terrestrial point cloud data using the DBSCAN clustering algorithm, it's necessary to first separate the non-terrestrial point cloud data into high-density and low-density data, and then cluster the data using the DBSCAN clustering algorithm.

[0024] It should be further explained that when dividing non-ground point cloud data into high-density and low-density data, ideally, high-density data should be surrounded by more high-density data, and low-density data should be surrounded by more low-density data. Therefore, multiple segmentation thresholds are used to divide non-ground point cloud data into high-density and low-density data. After each segmentation, the segmentation results at each segmentation threshold are determined based on whether each data point is of the same type as its nearest neighboring data. The optimal threshold is then determined, and the segmentation result at the optimal threshold is obtained.

[0025] Specifically, calculate the The distance between the first non-ground point cloud data and all non-ground point cloud data is obtained. The distance set of non-ground point cloud data is All data corresponding to the minimum distance in the distance set of non-ground point cloud data are recorded as The nearest neighbor data of non-ground point cloud data. Non-ground point cloud data and its Nearest neighbor data are used to obtain a nearest neighbor data pair. The Euclidean distance between the two data points in each nearest neighbor data pair is recorded as the nearest neighbor distance of the nearest neighbor data pair. A nearest neighbor distance set is obtained. The nearest neighbor distance set does not contain nearest neighbor data with the same value.

[0026] At this point, all the nearest neighbor data pairs and the nearest neighbor distance of each nearest neighbor data pair are obtained.

[0027] Furthermore, let the segmentation threshold be The non-ground point cloud data pairs within the nearest neighbor data pairs whose nearest neighbor distance is greater than the segmentation threshold are recorded as low-density data pairs; the non-ground point cloud data pairs within the nearest neighbor data pairs whose nearest neighbor distance is less than or equal to the segmentation threshold are recorded as high-density data pairs.

[0028] Further, the segmentation threshold is obtained as The segmentation effectiveness parameter when the neighbor distance is 0 is calculated as follows: ; Where, Indicates that the segmentation threshold is The segmentation effectiveness parameter when the neighbor distance is , represents the number of data points belonging to high-density data pairs, represents the number of data points belonging to low-density data pairs, For the The number of nearest neighbor data point pairs whose nearest neighbor distance is less than the segmentation threshold in all nearest neighbor data pairs of the data points belonging to the high-density data pairs; For the The number of nearest neighbor data point pairs whose nearest neighbor distances are greater than the segmentation threshold in all nearest neighbor data pairs of the data points belonging to the low-density data pair. Non-ground point cloud data belongs to both high-density data pairs and low-density data pairs, then calculate When The non-ground point cloud data are regarded as data points belonging to high-density data pairs, and the calculation When Non-ground point cloud data are regarded as data points belonging to low-density data pairs.

[0029] What needs to be explained is that The larger the value of The closer the neighbor distance is, the more consistent the characteristics of other non-ground point cloud data around the high-density data are. The greater the possibility that the distance between the nearest neighbors is the optimal segmentation threshold, The larger the value of is, the more consistent it is with the characteristics that other non-ground point cloud data around the low-density data are mostly low-density data, that is, The greater the probability that the distance between the two neighbors is the optimal segmentation threshold.

[0030] pass Calculate the formula to find When it is the maximum value, the distance value between the corresponding nearest neighbor point pairs is used later to divide the nearest neighbor data point pairs into high density and low density.

[0031] The ideal situation for high-density and low-density segmentation is that the nearest neighbor points of the high-density data point (the nearest neighbor point is the point with the closest Euclidean distance to the point, and there may be multiple) are all high-density data points, and the nearest neighbor points of the low-density data point are all low-density data points, so when When this reaches the maximum value, it means that more pairs of neighboring points belong to the same density category, which makes the division of high-density and low-density areas more coherent and reasonable.

[0032] Furthermore, the segmentation effectiveness parameter when the segmentation threshold is each nearest neighbor distance is calculated, and the nearest neighbor distance corresponding to the maximum segmentation effectiveness is recorded as the optimal segmentation threshold.

[0033] Furthermore, the non-ground point cloud data within the nearest neighbor data pair whose neighbor distance is greater than the optimal segmentation threshold is recorded as the final low-density data; the non-ground point cloud data within the nearest neighbor data pair whose neighbor distance is less than or equal to the optimal segmentation threshold is recorded as the final high-density data. The division results of high-density data pairs and low-density data pairs are as follows: Figure 10 In the figure, high-density data pairs are represented by blue, and high-density data pairs are connected by yellow straight lines; low-density data pairs are represented by red, and low-density data pairs are connected by black straight lines.

[0034] At this point, all non-ground data are divided into final high-density data and final low-density data.

[0035] Step S003: Use the non-dominated sorting genetic algorithm to obtain the DBSCAN clustering algorithm under the low-density parameter combination and the DBSCAN clustering algorithm under the high-density parameter combination in multiple parameter combinations. While ensuring that the core point is the final high-density data, use the DBSCAN clustering algorithm under the high-density parameter combination in each parameter combination to cluster all non-ground point cloud data to obtain noise points and high-density clustering clusters in each parameter combination; while ensuring that the core point is the final low-density data, use the DBSCAN clustering algorithm under the low-density parameter combination in each parameter combination to cluster the noise points in the parameter combination to obtain low-density clustering clusters and secondary noise points in each parameter combination; according to the distance between the secondary noise point and the low-density clustering cluster and the non-ground point cloud data in the high-density clustering cluster in each parameter combination, the final clustering cluster in each parameter combination is obtained.

[0036] It should be noted that when combining the non-dominated sorting genetic algorithm with the DBSCAN clustering algorithm, the non-dominated sorting genetic algorithm is first used to generate multiple parameter combinations for the neighborhood radius and minimum number of points within the DBSCAN clustering algorithm. The clustering results for each parameter combination are then obtained. The fitness of each clustering result is then determined using existing clustering result evaluation methods. Based on the fitness of the clustering results for each parameter combination, a relatively optimal parameter combination is then determined.

[0037] It should be further explained that the DBSCAN clustering algorithm is also called a density-based clustering algorithm, and its essence is to cluster data based on the number of other data around each data. When three-dimensional point cloud data is collected by a lidar, the amount of point cloud data collected for each object is related to the distance between the object and the lidar. Therefore, when the data is clustered by the DBSCAN clustering algorithm, some point cloud data of objects that are far away from the lidar will be regarded as noise points. Therefore, when the present invention clusters non-ground point cloud data by the DBSCAN clustering algorithm, all non-ground point cloud data are first clustered according to the neighborhood radius and the minimum number of points within a set of parameter combinations. Then, the DBSCAN clustering algorithm is used to cluster the noise points under the previous clustering result again, and the final clustering result is obtained based on the two clustering results.

[0038] It should be further explained that the secondary clustering of the noise points in the first clustering result is because when collecting 3D point cloud data using a lidar, the amount of 3D point cloud data of the collected object is inversely proportional to the distance between the lidars. That is, after the first clustering is completed, some 3D point cloud data of non-ground objects that are far away from the lidar may be regarded as noise points. Therefore, when re-clustering the noise points in the first clustering result, the core points can only be low-density data; and when performing the first clustering on all non-ground point cloud data, the core points of the clustering result can only be high-density data. When clustering data using the DBSCAN clustering algorithm, in order to avoid two clusters being mistakenly clustered into one class, the minimum neighborhood radius setting must be controlled. However, in this case, sparser points cannot be successfully clustered, causing some sparse points to be identified as noise points. Figure 7 is a schematic diagram of the distribution of noise points and data points within the cluster after clustering. Figure 7 In (a), the green points are a cluster, the blue points are a cluster, and the black points are noise points. However, these black points may not be noise points, but are formed due to the unreasonable setting of the neighborhood radius. Therefore, according to the distance between each noise point and the data points in other clusters around it, the noise points are redistributed, so that the data points mistakenly identified as noise points can be classified into clusters without changing the neighborhood radius. The image after the noise points are redistributed is shown in Figure 1. Figure 7 The clustering result image after redistribution of the noise points in (a) is Figure 7 Figure (b).

[0039] Specifically, a non-dominated sorting genetic algorithm is used to obtain multiple parameter combinations of neighborhood radius and minimum number of points. Using a non-dominated sorting genetic algorithm, multiple parameter combinations are obtained. Each parameter combination includes a high-density data parameter combination of neighborhood radius and minimum number of points, and a low-density parameter combination of neighborhood radius and minimum number of points.

[0040] Furthermore, the neighborhood radius and the minimum number of points under the high-density data parameter combination in the yth parameter combination are changed to the neighborhood radius and the minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the high-density parameter combination in the yth parameter combination is obtained.

[0041] Furthermore, the neighborhood radius and the minimum number of points under the low-density data parameter combination in the yth parameter combination are changed to the neighborhood radius and the minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the low-density parameter combination in the yth parameter combination is obtained.

[0042] Furthermore, while ensuring that the core points are the final high-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the high-density parameter combination within the y-th parameter combination is used to cluster all non-ground point cloud data, and the clustering clusters, core points and noise points under the high-density parameter combination within the y-th parameter combination are obtained, which are recorded as the high-density clustering clusters within the y-th parameter combination, the noise points and core points within this parameter combination.

[0043] Furthermore, while ensuring that the core points are the final low-density data, the DBSCAN clustering algorithm under the low-density parameter combination within the y-th parameter combination is used to continue clustering the noise points within the y-th parameter combination. The clustering clusters, noise points and core points obtained by this clustering are respectively recorded as the low-density clustering cluster within the y-th parameter combination, the secondary noise points within the y-th parameter combination and the core points within the y-th parameter combination.

[0044] Further, the yth parameter combination The Euclidean distance between the secondary noise point and each data in the initial cluster and the noise cluster is recorded as the yth parameter combination. The difference distance of the secondary noise points.

[0045] Furthermore, if the yth parameter combination If the minimum difference distance of the secondary noise points is less than or equal to the neighborhood radius of the high-density data parameter combination within the yth parameter combination, then the yth parameter combination within the yth parameter combination The secondary noise points are placed in the cluster corresponding to the data with the minimum difference distance.

[0046] Furthermore, the above operation is performed on all secondary noise points in the yth parameter combination to obtain the updated low-density clustering cluster and high-density clustering cluster in the yth parameter combination, and they are recorded as the final clustering cluster in the yth parameter combination.

[0047] Step S004: obtaining a first reference value and a second reference value of the final clustering result in each parameter combination according to the distribution of data around the core point in the final cluster in each parameter combination and the distribution of core points in different clusters.

[0048] It should be noted that the current evaluation of clustering results uses the distance from the data within a cluster to the cluster's centroid, as well as the distance between the centroids of different clusters, as a parameter for evaluating clustering results. However, this method is not suitable for judging the clustering results of density-based clustering algorithms. Figure 5 This is a schematic diagram of the results when clustering data using the DBSCAN clustering algorithm. For example, the accurate clustering results when clustering data using the DBSCAN clustering algorithm are as follows: Figure 5As shown in Figure (a), since the centroids of the two clusters are both in the center, the distance between the centroids of the two clusters is very close, so using traditional evaluation indicators to measure the clustering result is obviously not effective. Figure 5 In the case of (b), the centroids of the two clusters are on either side, so the distance between the two clusters is relatively far, and the distance between the points in a cluster and the centroid of the cluster is also close. Therefore, traditional evaluation indicators will consider this clustering result to be better. Therefore, the existing evaluation method of clustering results is improved. Figure 8 The result of clustering non-ground point cloud data using existing methods.

[0049] It should be further explained that when clustering data using the DBSCAN clustering algorithm, it is first determined whether the data is a core point based on whether the number of data within the neighborhood radius of each data is greater than the minimum number of points. Then, each core point and the data within its neighborhood radius are regarded as a cluster to complete the initial clustering. Afterwards, if a core point is within the cluster of another core point, the clusters where these two core points are located are regarded as the same cluster to complete the initial merging of clusters. That is, the essence of the DBSCAN clustering algorithm is to regard density-connected data as a cluster. Therefore, the mean of the Euclidean distance between a core point and all the data within its neighborhood radius is used as the density estimate of the core point. Then, the density consistency of the cluster is evaluated based on the variance of the density estimates of all core points in a cluster. That is, the present invention uses the sum of the density consistency of all clusters in the clustering result as an evaluation criterion for evaluating the clustering result.

[0050] It should be further explained that when the clustering effect of the cluster is judged by the variance of the density estimation value of the core points in all clusters at each clustering, it is found that when there is only one core point in each cluster, the variance of the density estimation value of the core points in the cluster is the smallest. In other words, when the clustering effect of the cluster is judged by the density variance of the core points in the cluster, the result of over-clustering may occur. Figure 6 This is a schematic diagram of over-clustering. Figure 6 Each different colored data point in the image represents a cluster, meaning that the data that should have been a single cluster has been mistakenly divided into multiple smaller clusters. These clusters are then separated by noise points. Therefore, during each clustering, the distance between the core points within each cluster is used to calculate the degree to which two clusters belong to the same cluster. This is then scored, and based on the distribution of noise points, a second reference value within the fitness of the clustering result is obtained.

[0051] It should be further explained that, when the data is clustered by the DBSCAN clustering algorithm, the core point is first obtained based on whether the number of data within the neighborhood radius of each data is greater than the minimum number of points. Afterwards, each core point and the data within its neighborhood radius are regarded as a cluster. That is, when the first clustering is completed by the DBSCAN clustering algorithm, each core point and all the data within its neighborhood radius are regarded as data within a cluster. Combined with the feature that a data cannot belong to two clusters at the same time, the present invention believes that if the Euclidean distance of the core points in two different clusters is less than or equal to 2 times the neighborhood radius, the two clusters are considered to be adjacent. Afterwards, the difference between the Euclidean distance between the core points in the two different clusters and the domain radius under the high-density parameter combination of 2 times is used as a parameter for calculating the segmentation score. The DBSCAN clustering algorithm is used to cluster data. In essence, the data is clustered according to the distribution of data around each data. That is, if the density variance of two clusters is closer, the possibility that the two clusters belong to the same cluster is greater. Therefore, the difference in the density variance of the two clusters is used as a parameter to calculate the segmented score.

[0052] It should be further explained that in real life, when clustering a set of data, it is less likely that there will be many noise points within the data set. Therefore, if the clustering result contains too many noise points, a larger weight is assigned to the over-segmentation score. If the clustering result contains too few noise points, a larger weight is assigned to the proportion of noise points and a smaller weight is assigned to the over-segmentation score, thus obtaining the second parameter for evaluating the effectiveness of the clustering result.

[0053] Specifically, it will be combined with the yth parameter The final cluster within the The data whose Euclidean distance of the core points is less than or equal to the neighborhood radius under the high-density parameter combination in the y-th parameter combination is recorded as the y-th parameter combination. The final cluster within the The neighborhood points of the core point.

[0054] Combine the yth parameter into the The final cluster within the The mean of the Euclidean distances between the core point and all its neighboring points is used as the The final cluster within the The density estimate of the core points.

[0055] Combine the yth parameter into the The variance of the density estimation value of all core points in the final cluster is recorded as the variance of the density estimation value of the yth parameter combination. The density consistency of the final clusters.

[0056] The sum of the density consistencies of all final clustering clusters in the yth parameter combination is recorded as the first reference value of the final clustering result in the yth parameter combination.

[0057] Further, calculate the yth parameter combination Each core point in the final cluster is The distance between each core point in the final cluster. The final cluster within the The first core point in the yth parameter combination The final cluster within the If the Euclidean distance between the core points of the two clusters is less than 2 times the radius of the high-density parameter combination under the yth parameter combination, the two clusters are considered adjacent. The difference between the Euclidean distance between the core points of the two different clusters and 2 times the radius of the high-density parameter combination under the yth parameter combination is used as a parameter for the over-segmentation score. The DBSCAN clustering algorithm is used to cluster data based on the distribution of the data surrounding each data point. That is, the closer the density variance of two clusters, the greater the likelihood that the two clusters belong to the same cluster. Therefore, the difference in the density variance of the two clusters is used as a parameter for calculating the over-segmentation score.

[0058] Further, get the yth parameter combination The final clusters and The specific calculation formula for the over-segmentation score of the final cluster is as follows: ; Where, The yth parameter combination The final clusters and The over-segmentation score of the final clusters, Indicates the yth parameter combination The variance of the density estimate of the core points in the final cluster, Indicates the yth parameter combination The variance of the density estimates of the cores within the final clusters, Indicates the yth parameter combination The final clusters and The number of over-subdivisions of the final clusters, Represents the neighborhood radius under the high-density data parameter combination within the yth parameter combination, Indicates the yth parameter combination The final clusters and The final cluster The degree of over-detailed performance, represents the absolute value function.

[0059] Where, Indicates the yth parameter combination The final clusters and The difference in the intra-cluster density variance between the final clusters is used to measure the The final clusters and The density similarity between the final clusters is consistent.

[0060] This is the result of the reduction. The original formula is The numerator represents the difference between the Euclidean distance between the core point a in the fth final cluster and the core point b in the dth final cluster and 2 times the neighborhood radius under the high-density data parameter combination in the yth parameter combination, because the maximum difference is 2 , the minimum is 0, so the meaning of the denominator here is to normalize. Finally, it is simplified to , is to calculate the similarity of the distance between two core points.

[0061] The degree of over-segmentation of the two clusters is measured by the distance between the core points. If there are multiple core points between the two clusters with a distance less than twice the neighborhood radius, then the possibility that the two clusters are actually one cluster is greater. It represents the difference in density consistency between the two clusters. That is to say, if the difference in density consistency between cluster A and cluster B is smaller, it means that the similarity of the intra-cluster density between cluster A and cluster B is higher.

[0062] In general, through We measure the similarity of density uniformity between any two clusters and the proximity of the core points within the cluster to determine whether the two clusters should be the same cluster but are over-segmented. The larger the value, the more serious the over-segmentation.

[0063] The sum of the over-segmentation scores of all two different final clusters in the yth parameter combination is input into Within the function, the output is recorded as the over-segmentation score of the final clustering result in the y-th parameter combination.

[0064] The following is a further explanation of the over-segmented scoring: Because some parameter combinations may result in clustering results that split data from a single cluster into several small, dense clusters, these clusters should actually belong to the same cluster. However, due to inappropriate parameter selection, these data are incorrectly clustered into multiple clusters. To avoid this, the present invention proposes an over-segmentation score as a reference value (as part of the second reference value). The over-segmentation score measures the degree to which data from a single cluster has been incorrectly clustered into different clusters.

[0065] Depend on Figure 5 (b) and Figure 6 It can be seen that these should have been clustered into one cluster, but due to the wrong selection of parameter combinations, the following is obtained: Figure 6 The clustering results shown are because the DBSCAN algorithm clusters all density-connected data, but it is obvious that Figure 6 These tiny clusters in the image are formed because they are over-segmented, so multiple small and dense clusters are formed. In other words, if the over-segmentation score is not added as a reference value, it is very likely that when the NSGAII algorithm automatically obtains parameters, the clustering results will be as follows: Figure 6 The parameter combinations shown.

[0066] By adding the over-segmentation score as a reference value, we can effectively avoid the situation where data from one cluster is incorrectly clustered into multiple clusters. This situation is called over-segmentation. The smaller the value of the over-segmentation score, the less severe the over-segmentation.

[0067] Furthermore, the calculation formula for obtaining the noise ratio of the final clustering result in the yth parameter combination is as follows: ; Where, Indicates the noise ratio of the final clustering result in the y-th parameter combination, Indicates the number of non-ground point cloud data, Indicates the number of non-ground point cloud data that are not in the final cluster within the y-th parameter combination.

[0068] Furthermore, the specific calculation formula for obtaining the second reference value within the fitness of the clustering result under the high-density parameter combination within the y-th parameter combination is as follows: ; Where, The second parameter representing the final clustering result in the y-th parameter combination, represents the over-segmentation score of the final clustering result in the y-th parameter combination, Indicates the noise proportion of the final clustering result within the yth parameter combination." ” represents the proportion of non-noise data points, Indicates the proportion of noise, here we add This is to control the proportion of noise. If the proportion of noise is too large, then The value will also increase accordingly, which can effectively avoid the occurrence of a parameter combination that makes the noise ratio in the clustering result very large.

[0069] Step S005: The first reference value and the second reference value of the final clustering result in each parameter combination are input into the non-dominated sorting genetic algorithm as two fitnesses to obtain the optimal parameter combination when clustering the non-ground point cloud data using the DBSCAN clustering algorithm; all non-ground point cloud data are clustered multiple times by using the DBSCAN clustering algorithm under the optimal parameter combination when clustering the non-ground point cloud data using the DBSCAN clustering algorithm to obtain the final clustering result, thereby completing the segmentation of the non-ground point cloud data.

[0070] Specifically, the first reference value and the second reference value of the final clustering result in the high-density parameter combination in each parameter combination are input into the non-dominated sorting genetic algorithm as the two fitness of the parameter combination to obtain the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm.

[0071] The neighborhood radius and minimum number of points in the optimal parameter combination when using the DBSCAN clustering algorithm to cluster non-ground point cloud data are used as the neighborhood radius and minimum number of points within the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the high-density parameter combination within the optimal parameter combination is obtained.

[0072] While ensuring that the core points are the final high-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the high-density parameter combination within the optimal parameter combination is used to cluster all non-ground point cloud data. The cluster clusters and noise points under the high-density parameter combination within the optimal parameter combination are obtained, which are recorded as the initial cluster clusters and initial noise points respectively.

[0073] The neighborhood radius and the minimum number of points under the low-density data parameter combination within the optimal parameter combination when clustering all non-ground point cloud data are set as the neighborhood radius and the minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the low-density parameter combination within the optimal parameter combination is obtained.

[0074] While ensuring that the core points are the final low-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the low-density parameter combination within the optimal parameter combination is used to cluster all non-ground point cloud data. The cluster clusters and noise points under the low-density parameter combination within the optimal parameter combination are obtained, which are recorded as noise cluster clusters and secondary noise respectively.

[0075] Further, the The Euclidean distance between the secondary noise point and each data in the initial cluster and the noise cluster is recorded as The difference distance of the secondary noise points. The minimum difference distance of the secondary noise points is less than or equal to the optimal neighborhood when clustering the initial noise points, then the Each secondary noise point is placed in the cluster corresponding to the data with the minimum difference distance. The above operation is repeated for all secondary noise points to obtain updated initial clusters and updated noise clusters. Based on the updated initial clusters and noise clusters, the final clustering results are obtained. The final clustering results complete the segmentation of the non-ground point cloud data. Figure 9 This is the result of clustering non-ground point cloud data using the method of the present invention.

[0076] This method uses secondary clustering and distance verification to prevent low-density objects from being misclassified as noise, reducing missed classifications. It also eliminates true noise, lowering the misclassification rate and ensuring that the segmentation results of non-ground point clouds more closely resemble the distribution of objects in real scenes (for example, objects of varying densities, such as nearby vehicles and distant signboards, can all be correctly classified). By first partitioning, then performing targeted clustering, and finally verifying the optimization process, this method effectively overcomes the impact of uneven point cloud density on DBSCAN clustering and improves the accuracy of non-ground point cloud segmentation.

[0077] Another embodiment of the present invention provides a point cloud data segmentation system based on the DBSCAN clustering algorithm, comprising: Point cloud data acquisition module, used to obtain non-ground point cloud data.

[0078] The point cloud data segmentation module is used to divide the non-ground point cloud data into final high-density data and final low-density data according to the distribution of the non-ground point cloud data.

[0079] While ensuring that the core points are the final high-density data, the DBSCAN clustering algorithm is used to cluster all non-ground point cloud data to obtain initial clusters and initial noise points.

[0080] While ensuring that the core points are the final low-density data, the DBSCAN clustering algorithm is used to cluster the initial noise points to obtain noise clusters and secondary noise points.

[0081] According to the Euclidean distance between the secondary noise points and the initial clusters and each data in the noise clusters, the final clustering result is obtained, and the segmentation of the non-ground point cloud data is completed.

[0082] It should be noted that the specific embodiments described above can enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although this specification and examples have described the present invention in detail, those skilled in the art should understand that the present invention can still be modified or replaced with equivalents; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered by the scope of protection of the patent for the present invention.

Claims

1. A point cloud data segmentation method based on DBSCAN clustering algorithm, characterized in that: include: Acquire non-ground point cloud data; According to the distribution of non-ground point cloud data, the non-ground point cloud data is divided into final high-density data and final low-density data; While ensuring that the core points are the final high-density data, the DBSCAN clustering algorithm is used to cluster all non-ground point cloud data to obtain the initial clusters and initial noise points; While ensuring that the core points are the final low-density data, the DBSCAN clustering algorithm is used to cluster the initial noise points to obtain noise clusters and secondary noise points; According to the Euclidean distance between the secondary noise points and the initial clusters and each data in the noise clusters, the final clustering result is obtained, and the segmentation of the non-ground point cloud data is completed.

2. The point cloud data segmentation method based on the DBSCAN clustering algorithm according to claim 1, characterized in that: The non-ground point cloud data is divided into final high-density data and final low-density data according to the distribution of the non-ground point cloud data. The specific steps are as follows: Calculate the The Euclidean distance between the first non-ground point cloud data and all non-ground point cloud data is obtained. A distance set of non-ground point cloud data; The first All data corresponding to the minimum distance in the distance set of non-ground point cloud data are recorded as The nearest neighbor data of non-ground point cloud data; Through the Non-ground point cloud data and its nearest neighbor data, and get a nearest neighbor data pair; The Euclidean distance between the two data in each nearest neighbor data pair is recorded as the nearest neighbor distance of the nearest neighbor data pair, and the nearest neighbor distance set is obtained; Each neighbor distance in the neighbor distance set is used as a segmentation threshold, and high-density data and low-density data under each segmentation threshold are obtained according to the segmentation threshold; According to the high-density data and low-density data in each neighbor data pair under each segmentation threshold, the final high-density data and the final low-density data are obtained.

3. The point cloud data segmentation method based on the DBSCAN clustering algorithm according to claim 2, characterized in that: The specific steps of obtaining high-density data and low-density data under each segmentation threshold according to the segmentation threshold are as follows: Let the segmentation threshold be The non-ground point cloud data pairs within the nearest neighbor data pairs whose nearest neighbor distance is greater than the segmentation threshold are recorded as low-density data pairs; The non-ground point cloud data pairs within the nearest neighbor data pairs whose neighbor distance is less than or equal to the segmentation threshold are recorded as high-density data pairs.

4. The point cloud data segmentation method based on the DBSCAN clustering algorithm according to claim 2, characterized in that: The specific steps of obtaining the final high-density data and the final low-density data according to the high-density data and the low-density data in each neighboring data pair under each segmentation threshold are as follows: Get the segmentation threshold as The segmentation effectiveness parameter when the neighbor distance is 0 is calculated as follows: ； Where, Indicates that the segmentation threshold is The segmentation effectiveness parameter when the neighbor distance is , represents the number of data points belonging to high-density data pairs, represents the number of data points belonging to low-density data pairs, For the The number of nearest neighbor data point pairs whose nearest neighbor distance is less than the segmentation threshold in all nearest neighbor data pairs of the data points belonging to the high-density data pairs; For the The number of nearest neighbor data point pairs whose nearest neighbor distance is greater than the segmentation threshold in all nearest neighbor data pairs of the data points belonging to the low-density data pairs; The nearest neighbor distance corresponding to the maximum segmentation effectiveness is recorded as the optimal segmentation threshold; The non-ground point cloud data within the nearest neighbor data pair whose neighbor distance is greater than the optimal segmentation threshold is recorded as the final low-density data; the non-ground point cloud data within the nearest neighbor data pair whose neighbor distance is less than or equal to the optimal segmentation threshold is recorded as the final high-density data.

5. The point cloud data segmentation method based on the DBSCAN clustering algorithm according to claim 1, characterized in that: It also includes combining the non-dominated sorting genetic algorithm with the DBSCAN clustering algorithm to obtain the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm, and obtaining the final segmentation result by using the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm.

6. The point cloud data segmentation method based on the DBSCAN clustering algorithm according to claim 5, characterized in that: The optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm obtains the final segmentation result. The specific steps are as follows: A non-dominated sorting genetic algorithm is used to obtain multiple parameter combinations; each parameter combination includes a high-density data parameter combination of neighborhood radius and minimum number of points, and a low-density parameter combination of neighborhood radius and minimum number of points, and the neighborhood radius and minimum number of points in the high-density parameter combination are different from those in the low-density parameter combination; The neighborhood radius and minimum number of points under the high-density data parameter combination in the y-th parameter combination are used as the neighborhood radius and minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the high-density parameter combination in the y-th parameter combination is obtained; The neighborhood radius and minimum number of points under the low-density data parameter combination in the y-th parameter combination are used as the neighborhood radius and minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the low-density parameter combination in the y-th parameter combination is obtained; Under the condition that the core point is the final high-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the high-density parameter combination within the y-th parameter combination is used to cluster all non-ground point cloud data, and the cluster clusters, core points and noise points under the high-density parameter combination within the y-th parameter combination are obtained, which are recorded as the high-density cluster cluster within the y-th parameter combination, the noise points and core points within the parameter combination; Under the condition that the core points are the final low-density data, the DBSCAN clustering algorithm under the low-density parameter combination in the y-th parameter combination is used to continue clustering the noise points in the y-th parameter combination. The clusters, noise points and core points obtained by this clustering are recorded as the low-density clusters in the y-th parameter combination, the secondary noise points in the y-th parameter combination and the core points in the y-th parameter combination respectively. Combine the yth parameter into the The Euclidean distance between the secondary noise point and each data in the initial cluster and the noise cluster is recorded as the yth parameter combination. The difference distance of the secondary noise points; If the yth parameter combination If the minimum difference distance of the secondary noise points is less than or equal to the neighborhood radius of the high-density data parameter combination within the yth parameter combination, then the yth parameter combination within the yth parameter combination The secondary noise points are placed in the cluster corresponding to the data with the minimum difference distance; The above operation is performed on all secondary noise points in the yth parameter combination to obtain the updated low-density clusters and high-density clusters in the yth parameter combination, and they are recorded as the final clusters in the yth parameter combination; Obtaining a first reference value of the final clustering result in the yth parameter combination through the distribution of the core point in the final cluster and the data around the core point in the yth parameter combination; According to the distances of the core points in different final clustering clusters in the yth parameter combination, the distribution differences of the data around the core points, and the distribution of the noise points, a second reference value of the final clustering result in the yth parameter combination is obtained; The first reference value and the second reference value of the final clustering result in each parameter combination are input into the non-dominated sorting genetic algorithm as the two fitness values ​​of the parameter combination to obtain the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm; The neighborhood radius and the minimum number of points in the optimal parameter combination when clustering non-ground point cloud data using the DBSCAN clustering algorithm are used as the neighborhood radius and the minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the high-density parameter combination in the optimal parameter combination is obtained; While ensuring that the core points are the final high-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the high-density parameter combination within the optimal parameter combination is used to cluster all non-ground point cloud data. The cluster clusters and noise points under the high-density parameter combination within the optimal parameter combination are obtained, which are recorded as the initial cluster clusters and initial noise points respectively. The neighborhood radius and minimum number of points under the low-density data parameter combination within the optimal parameter combination when clustering all non-ground point cloud data are set as the neighborhood radius and minimum number of points in the DBSCAN clustering algorithm, and the DBSCAN clustering algorithm under the low-density parameter combination within the optimal parameter combination is obtained; While ensuring that the core points are the final low-density data, the Euclidean distance is used as the distance metric, and the DBSCAN clustering algorithm under the low-density parameter combination within the optimal parameter combination is used to cluster all non-ground point cloud data. The cluster clusters and noise points under the low-density parameter combination within the optimal parameter combination are obtained, which are recorded as noise cluster clusters and secondary noise points respectively. The final clustering result is obtained based on the Euclidean distance between the secondary noise point and the initial cluster and each data in the noise cluster.

7. The point cloud data segmentation method based on the DBSCAN clustering algorithm according to claim 6, characterized in that: The first reference value of the final clustering result in the yth parameter combination is obtained by the distribution of the core point in the final cluster and the data around the core point in the yth parameter combination. The specific steps are as follows: Combined with the yth parameter The final cluster within the The data whose Euclidean distance of the core points is less than or equal to the neighborhood radius under the high-density parameter combination in the y-th parameter combination is recorded as the y-th parameter combination. The final cluster within the Neighborhood points of the core point; Combine the yth parameter into the The final cluster within the The mean of the Euclidean distances between the core point and all its neighboring points is used as the The final cluster within the The density estimate of the core points; Combine the yth parameter into the The variance of the density estimation value of all core points in the final cluster is recorded as the variance of the density estimation value of the yth parameter combination. The density consistency of the final clusters; The sum of the density consistencies of all final clustering clusters in the yth parameter combination is recorded as the first reference value of the final clustering result in the yth parameter combination.

8. The point cloud data segmentation method based on the DBSCAN clustering algorithm according to claim 7, characterized in that: The second reference value of the final clustering result in the yth parameter combination is obtained according to the distances between the core points in different final clustering clusters in the yth parameter combination, the distribution differences of the data around the core points, and the distribution of the noise points. The specific steps are as follows: Calculate the yth parameter combination Each core point in the final cluster is The distance between each core point in the final cluster; If the yth parameter combination The final cluster within the The first core point in the yth parameter combination The final cluster within the If the Euclidean distance between the core points is less than 2 times the neighborhood radius under the high-density parameter combination in the y-th parameter combination, the two clusters are considered adjacent. Then the difference between the Euclidean distance between the core points in the two different clusters and 2 times the neighborhood radius under the high-density parameter combination in the y-th parameter combination is used as a parameter for the over-segmentation score; Get the yth parameter combination The final clusters and The specific calculation formula for the over-segmentation score of the final cluster is as follows: ； Where, The yth parameter combination The final clusters and The over-segmentation score of the final clusters, Indicates the yth parameter combination The variance of the density estimate of the core points in the final cluster, Indicates the yth parameter combination The variance of the density estimates of the cores within the final clusters, Indicates the yth parameter combination The final clusters and The number of over-subdivisions of the final clusters, Represents the neighborhood radius under the high-density data parameter combination within the yth parameter combination, Indicates the yth parameter combination The final clusters and The final cluster The degree of over-detailed performance, represents the absolute value function; The sum of the over-segmentation scores of all two different final clusters in the yth parameter combination is input into the normalization function, and the output is recorded as the over-segmentation score of the final clustering result in the yth parameter combination; The calculation formula for obtaining the noise ratio of the final clustering result in the yth parameter combination is as follows: ； Where, Indicates the noise ratio of the final clustering result in the y-th parameter combination, Indicates the number of non-ground point cloud data, Indicates the number of non-ground point cloud data that are not in the final cluster within the y-th parameter combination; The specific calculation formula for obtaining the second reference value within the fitness of the clustering result under the high-density parameter combination within the y-th parameter combination is as follows: ； Where, The second parameter representing the final clustering result in the y-th parameter combination, represents the over-segmentation score of the final clustering result in the y-th parameter combination, Indicates the noise proportion of the final clustering result within the y-th parameter combination.

9. The point cloud data segmentation method based on the DBSCAN clustering algorithm according to claim 1, characterized in that: The final clustering result is obtained based on the Euclidean distance between the secondary noise point and the initial cluster and each data in the noise cluster. The specific steps are as follows: The first The Euclidean distance between the secondary noise point and each data in the initial cluster and the noise cluster is recorded as The difference distance of the secondary noise points; Jordi If the minimum difference distance of the secondary noise points is less than or equal to the neighborhood radius of the DBSCAN clustering algorithm under the optimal high-density parameter combination when clustering all non-ground point cloud data, then the The secondary noise points are placed in the cluster corresponding to the data with the minimum difference distance; The above operation is performed on all secondary noise points to obtain updated initial clusters and updated noise clusters; the final clustering results are obtained based on the updated initial clusters and noise clusters.

10. A point cloud data segmentation system based on DBSCAN clustering algorithm, characterized in that: include: Point cloud data acquisition module, used to obtain non-ground point cloud data; A point cloud data segmentation module is used to divide the non-ground point cloud data into final high-density data and final low-density data according to the distribution of the non-ground point cloud data; While ensuring that the core points are the final high-density data, the DBSCAN clustering algorithm is used to cluster all non-ground point cloud data to obtain the initial clusters and initial noise points; While ensuring that the core points are the final low-density data, the DBSCAN clustering algorithm is used to cluster the initial noise points to obtain noise clusters and secondary noise points; According to the Euclidean distance between the secondary noise points and the initial clusters and each data in the noise clusters, the final clustering result is obtained, and the segmentation of the non-ground point cloud data is completed.

Images