Classification Model of 3D Point Cloud Based on Linked Adaptive Graph Convolution

Abstract

A three-dimensional point cloud classification model LAGCN based on adaptive graph convolution is proposed to solve the problem that PointNet only extracts feature information from isolated points but not from neighborhood structure. Firstly, the three-dimensional point cloud is transformed into undirected graph structure, which is used to obtain the neighborhood information of point cloud adaptively, and the classification accuracy is improved by incorporating residual structure. In the classification experiment, this paper trains and tests on ModelNet40 dataset, and compares the classification accuracy with VoxNet, PointNet and DGCNN models. The classification accuracy of this model is better than that of the above models.