Point Clouds Voxel Oversegmentation Method Based On High-Performance Search In Three-Dimensional Space

Lidar recently is a high-tech to obtain 3D point clouds scenes.The point clouds data obtained by high-precision lidar possess rich geometric information of surrounding environment.However,due to the large amount of 3D point clouds data,there will be inevitable noise interference when the lidar scans the environment.Meanwhile,the point clouds have a certain uneven density distribution.It is a challenging task to quickly and accurately enhance and oversegment the road infrastructure of point cloud scanned by mobile laser system and terrestrial laser system.Therefore,this thesis utilizes the nearest neighbor search methods and spatial judgment conditions based on octree uniform storage point cloud data with height and angle to oversegment the point clouds data.The research contents are presented as follows:Firstly,an improved octree point cloud over segmentation framework based on neighborhood information is constructed.This thesis uses three different search methods to obtain the surrounding points of each point stored in the octree,so as to consider the corresponding nearest neighbor normal vector information.The local k-means clustering method is used to cluster the three kinds of nearest neighbor information.The clustering points are labeled to get the corresponding over segmentation results.In this thesis,the effectiveness of the proposed oversegmentation framework is evaluated by applying the public datasets obtained from terrestrial laser scanning and mobile laser scanning and the commonly used metric evaluation methods of over segmentation algorithm.The results show that the boundary recall rate of the proposed method is about 7 and 4 times higher than that of the traditional VCCS method,and the running time is feasible and effective.Secondly,the oversegmentation algorithm of 3D point clouds for road boundaries enhancement is constructed.First,this paper uses the radius k nearest neighbor search method to segment seed points on octrees to obtain the neighborhood information.Second,the iterative weighted least square method and the spatial structure judgment method based on angle and height threshold are used to oversegment point clouds depending on seed points.Finally,supervoxel boundaries are fine tuned by updating supervoxel centroids with the neighbor information.In this paper,two large outdoor point cloud benchmarks,IQmulus and TerraMobilita and Semantic 3D dataset,are utilized to demonstrate the effectiveness of the algorithm.Experimental results show that the proposed oversegmentation algorithms are better than other contrast methods in road boundaries enhancement tasks.