Research On Point Cloud Semantic Classification And Segmentation Model Based On Deep Learning

Because 3D point cloud data contains rich semantic information,and has the characteristics of high density,high accuracy,and massive data,it has gradually become one of the main data for understanding and analyzing natural scenes.At the same time,because of the huge research value of 3D point cloud data,it has quickly become a research hotspot in the field of computer vision,and has been widely used in artificial intelligence,aerospace,intelligent transportation and other fields.In the study of semantic classification segmentation,three-dimensional point cloud is not affected by environmental factors such as illumination and occlusion,and contains depth features,which can well show the spatial information of complex scene.Therefore,the research of point cloud semantic classification segmentation has a great impetus for the application of computer vision field.The semantic classification segmentation of 3D point cloud data is closely related to the application of point cloud data set and depth learning technology in 3D data.At present,the research on 3D point cloud data has made great progress,but the recognition accuracy of existing deep learning model point cloud semantic classification and segmentation methods still has much room for improvement.On the one hand,the irregularity and disorder of point cloud data make it impossible for the existing neural network model to perform convolution operation on the data.on the other hand,there is no suitable feature description method to extract local and global features of point cloud under different scenes.In this paper,the above problems in the classification and segmentation of 3D point cloud semantics are deeply studied.The specific tasks are as follows:(1)Aiming at the problem that two-dimensional convolution cannot be applied to disordered three-dimensional point clouds,a segmentation network Spider-Net that directly performs convolution operations on point cloud data is proposed.By introducing a parameterized convolution module SpiderCNN that combines a step function and a third-order Taylor,the geodesic distance of adjacent points in the local area of the point cloud is used as the convolution object to perform the convolution operation,and then the network can directly extract the features of the irregular point cloud data.The network model achieved an overall segmentation accuracy of 85.3%and an average intersection ratio(MIOU)of 60.1% on the indoor segmentation data set S3DIS.(2)The classification and segmentation accuracy of point cloud data is closely related to the ability of the network to describe global and local features.a point cloud semantic classification segmentation network that combines GCN and differential pooling functions is proposed.By introducing a graph convolutional neural network into the classification segmentation model to continue convolution of the edges of adjacent point pairs in the graph structure to obtain local features,Meanwhile,different pooling functions are used to extract multiple global features and obtain more abundant feature information.The model achieved an overall classification accuracy of 92.1% on ModelNet40,and the average intersection ratios on the ShapeNet and S3DIS datasets were 86.1% and 57.4%,respectively.Comparing the point cloud semantic classification and segmentation model proposed in this paper with other existing research methods,the research method proposed in this paper achieves a certain improvement in the accuracy of point cloud classification and segmentation.