Research On Semantic Segmentation Of 3D Point Cloud

In recent years,unmanned driving systems have entered a practical stage as an important application field of artificial intelligence.The unmanned driving system detects and perceives the surrounding environment and its various states through on-board sensors,and then recognizes the surrounding environment through relevant algorithms and assigns classification information for the system to make analysis and decision.Among them,lidar is a kind of sensor with application prospect in this field,so it is very important to study its point cloud processing method.Based on the semantic segmentation of 3D point clouds,this article is mainly divided into the following two aspects:(1)Adopt traditional point cloud segmentation method: first perform ground filtering,and then use the kmeans and spectral clustering and other clustering algorithms to perform clustering on the remaining point cloud to obtain segmentation results.Among them,kmeans clustering depends on the selection of the number of classification categories and the selection of the center point of initialization,and is sensitive to noise points;the clustering result of meanshift clustering algorithm depends on the bandwidth setting,the bandwidth setting is too small,the convergence is too slow,the cluster There are too many classes,the bandwidth setting is too large,and some clusters may be lost;the selection of related parameters by the spectral clustering algorithm directly affects the generation of the adjacency matrix and the number of clusters,but the spectral clustering can be in any shape The advantage of clustering results in the global data set is that the clustering results of DBCSAN are sensitive to the radius of the selected neighborhood and the number of point clouds in the neighborhood,but insensitive to the shape of the dataset space.Through the analysis of the clustering process and results of each method,it is found that the effect of the traditional method will depend on the relevant parameters manually set,and the generalization ability is poor.This is inefficient for unmanned driving systems to automatically adapt to complex environments,and poor generalization ability.(2)Neural network for point cloud segmentation: In this paper,Semantic3 D data set is used.This data set is collected by a static laser scanner placed on the ground,and is divided into a training set and a test set.Drawing on the structure of pointnet using shared mlp and maxpooling to solve the point cloud disorder feature,the relevant parameters are optimized according to the different inputs of the data set,and the point cloud rotation part of the network structure is simplified according to the characteristics of the data set,Built a benchmark network.Then according to the problem of pointnet's lack of extraction of the internal information of the point cloud when processing the point cloud,the kmeans clustering method and meanshift method are added to read the local information of the point cloud to improve,and the local information obtained by clustering and the reference network The global information is combined and classified,and the Semantic3 d data set is used for experiments.Compared with the benchmark network,the overall average accuracy is increased from 74% to 79% and 85%,The classification accuracy rate and recall rate of most categories have improved.