Research Of LIDAR Point Cloud Feature Extraction And Segmentation

3D point cloud has the characteristics of high density and precision,and its coordinate information contains a lot of semantic features.At present,the semantic segmentation of the 3D point cloud is an effective way to solve the problem of self-driving and build a high-precision map.Currently,3D semantic segmentation is mainly based on point cloud data,suffering large amounts,uneven density and irregularity.The feature extraction model based on KNN is an advanced method of the point cloud learning.However,the feature extraction model based on KNN has some shortcomings in the calculation of point cloud neighborhood and the training process.To solve the above problems,this paper aim at the features extraction of point cloud and the accuracy of point cloud segmentation.The specific work of this paper is as follows:(1)this paper proposes a point cloud semantic segmentation algorithm based on MSKNN neighborhood search,This algorithm mainly solves the shortage of KNN neighbor feature extraction model in the density of point cloud and improves the accuracy of point cloud semantic segmentation by using Mean-shift and KNN methods to calculate the neighbor and extract features.The algorithm consists of three steps.First,the Mean-shift and KNN methods are used to calculate the nearest neighbors of points.Based on different point cloud density,different point cloud nearest neighbors are calculated.Secondly,the geometric features of the point cloud are extracted from different neighbors,and then the corresponding labels are assigned to each point through the classifier.Thirdly,we use a plane divided power line category to fuse the classification labels based on Mean-shift and KNN to get the final semantic labels.(2)This paper proposes a multiscale point cloud semantic segmentation algorithm based on downsampling.The main purpose of this algorithm is to solve the problem of calculating and optimizing the nearest neighbor for each point when calculating the nearest neighbor of the point cloud.Under the condition of ensuring the segmentation accuracy,it can avoid the query of the nearest neighbor of a large number of points and accelerate the training and prediction.Firstly,the point cloud is downsampled to get different density point clouds.Secondly,the ball search is used to calculate the multi-scale nearest neighbor of the point cloud.Finally,the geometric features of the point cloud are calculated,and the final semantic label is obtained by using the classifier.(3)This paper proposes an improved regularization framework algorithm,which is mainly to solve the problem of noise points in the classified point cloud.By regularizing the point cloud,the noise points in the point cloud label are filtered out,the segmentation accuracy is improved,and the label has spatial smoothness.Firstly,the point cloud adjacency graph is selected,the penalty function is constructed,and the optimization algorithm is used to get the final semantic labels.The proposed method is tested and analyzed in Oakland and self-built datasets in detail and compared with the KNN neighborhood feature extraction model in many aspects.The experimental results show that the overall accuracy of the MS-KNN neighborhood search point cloud semantic segmentation algorithm is increased by 2.24%,and the average of F1 increased by 7.83%.A multiscale point cloud semantic segmentation algorithm based on downsampling reduces the overall accuracy by 1.86% but reduces the training time by 13 times.The overall accuracy of the regularization framework algorithm is improved by 3.53%,and the designed exponential fidelity function is improved by 0.2%compared with other fidelity functions.