Research On Road Scene Classification Based On LiDAR Point Cloud Dat

In the geospatial information science oriented technology,Li DAR(Laser radar)has become an important method to acquire three-dimensional spatial location information and attribute information.Nowadays,3D point cloud has become the third kind of spatial data besides maps and images,and plays a very key role in digital earth,digital city,intelligent transportation and 3D mapping.In the process of point cloud data processing,fast,efficient and high-precision automatic point cloud segmentation and classification technology is the key to realize the effective utilization of point cloud data.The technical difficulties of point cloud segmentation and classification mainly exist in two aspects: one is the huge amount of point cloud data,uneven density,irregular structure,mutual occlusion between ground objects,data redundancy and data loss coexist;Second,the existing relatively advanced classification methods are generally not robust,and the application of various classification algorithms in large-scale scenarios has obvious limitations.The above difficulties are particularly prominent in the point cloud classification of vehicle-mounted Li DAR.In order to achieve fast and accurate classification of ground objects in geographical complex scenes,especially on both sides of roads,a point cloud classification method combining point cloud segmentation and improved adaptive neighborhood feature extraction is proposed in this paper.The obtained point cloud data are mainly divided into buildings,ground points,vegetation,lampposts and other ground objects.The main work of this paper is as follows:(1)Aiming at the shortcomings of the current 3D point cloud data segmentation,such as poor robustness and low efficiency,this paper proposes a segmentation algorithm combining hypervoxel and P-Linkage clustering to achieve accurate and fast segmentation of 3D point cloud data from top to bottom.Firstly,the point cloud is generated into a hypervoxel which can retain the complete boundary information,and then the segmentation region with high precision and high boundary recall rate is obtained by linking clustering on the basis of the hypervoxel,which lays a good foundation for correcting the boundary errors of classification.(2)To solve the problem of insufficient and inaccurate neighborhood feature extraction,an improved adaptive neighborhood selection algorithm is proposed in this paper.Based on the existing feature entropy function to determine the range of neighborhood,a distance weighting function is introduced in the algorithm,which makes the point closer to the search point be given a higher weight and get more accurate neighborhood information,so as to accurately extract the feature information of the search point and improve the accuracy of classification.(3)Aiming at the problem that the classification boundary is prone to error due to the complex structure of ground objects,the serious occlusion between ground objects and the uneven density of point cloud,this paper proposes an algorithm to correct the classification boundary by using high-precision region segmentation.The algorithm projected the segmented region and the initial classification results in the same coordinate system and corrected the classification results with the boundary of the segmented region as the standard so as to get the correct boundary classification point cloud.The algorithm improves the overall accuracy of final classification to a certain extent.