A Point Cloud Classification Model Based On Improved Dynamic Graph Convolution

With the rapid development of 3D point cloud acquisition technology,point cloud data has become a research hotspot in the fields of autopilot,virtual reality,augmented reality and robot.As an important branch of point cloud data processing,point cloud classification task has important research significance and research value.In view of the sparsity,irregularity and disorder of point cloud,this is a very challenging task.In most methods to deal with this problem,the local feature information extraction of point cloud is insufficient,and the accuracy and robustness of classification are not high.This paper studies the above problems and gives a new improvement method.Firstly,aiming at the problem of insufficient local feature extraction in dynamic graph convolution,based on the point cloud dynamic graph convolution,a dynamic graph convolution method weighted by position relationship is proposed.The position relationship of feature vector is learned by sharing multi-layer perceptron,the weight coefficient matrix is generated,and the dynamic graph convolution results are dynamically weighted.The improved convolution method is used to construct the classification network model.Secondly,the influence of different spatial position relationship expressions of feature vectors on the ability of local feature extraction is studied,and three different position relationship expressions are designed to obtain the volume weight coefficient matrix of dynamic map.Aiming at the problem of the influence of the number of improved dynamic graph layers on the network performance,the improved dynamic graph layers are gradually increased from less to more,and the influence of the number of improved dynamic graph layers on the performance of the network model is studied.In view of the serious loss of information in the process of global pooling,the maximum pooling splicing global pooling is used to extract the global feature information.Then,aiming at the problem of low performance of sparse point cloud model,the network structure of point cloud classification is optimized combined with the idea of dense connection.The edge vector with rich information is calculated by connecting the hierarchical features of different dynamic graphs,which reduces the problem of gradient disappearance,and effectively improve the performance of sparse point cloud classification.Finally,the experimental verification of the dynamic graph convolution point cloud classification model based on position relationship weighting proposed in this paper is carried out based on the standard public data set modelnet40.The experimental results show that the proposed method can make a good classification effectly.