Research On 3D Point Cloud Classification And Segmentation Base On Deep Learning

As 3D point cloud data is widely used in the fields of robots,unmanned driving,and3 D scene roaming,the amount of 3D point cloud data has increased dramatically,so it has important practical significance to propose an efficient and intelligent 3D target recognition method.At present,deep learning has made great progress in the field of computer vision.This article applies deep learning to the 3D target recognition task and obtains the following research results and conclusions:3D point cloud classification and segmentation network based on Spider convolution.Aiming at the problem that Spider Convolutional Neural Network(CNN)cannot fully obtain deep-level feature information from 3D point cloud data,a classification and segmentation network that can directly handle 3D point cloud data is designed:Linked-Spider CNN,this method is based on the combination of Spider CNN and residual network idea.First of all,by increasing the number of Spider convolution layers on the basis of Spider CNN to obtain deep-level features of the point cloud;secondly,the idea of introducing a residual network adds short connections to each layer of Spider convolution to form a residual block,and convolves each layer of Spider The output is stitched and fused to form point cloud feature data,and then the point cloud feature data is used for classification and segmentation.Experimental results show that compared with Point Net++,Spider CNN and other networks,the proposed network can improve the classification accuracy and segmentation effect of point clouds,and prove that the network has faster convergence speed and stronger robustness.A point cloud classification network based on the combination of global and local correlation information.Although Linked-Spider CNN is outstanding in point cloud recognition tasks,it does not consider the high-order correlation information between point cloud feature data.To solve this problem,this paper designs a point cloud classification network based on the combination of global and local correlation information on the basis of Hypergraph Neural Networks(HGNN)combined with hypergraph random walk.First,the point cloud feature data is constructed as a hypergraph,and the global correlation information between the point cloud feature data is obtained through the random walk process on the hypergraph,and the local correlation information between the point cloud feature data is obtained using hyperedge convolution.Finally,the local and global correlation information is stitched and fused,and input to the classifier.Experiments prove that the proposed network has better classification performance than HGNN and other networks.