A Point Cloud Registration Network Combined with Graph Convolution and FPFH for Feature Extraction

Abstract

Currently most point cloud registration networks only take the 3D coordinates of the point cloud as the network inputs, then map the point cloud into a high-dimensional feature space by using the neural network model, and use the features from the high-dimensional feature space for point cloud registration. However the geometric prior knowledge of the point cloud structure is not fully used during the registration process. Therefore, a novel point cloud registration network is proposed in this paper. In the proposed network model, a Graph Convolutional Network (GNN) for feature extraction is applied, meanwhile the traditional Fast Point Feature Histograms (FPFH) feature extraction method is also applied. The geometric features acquired by FPFH and the high dimensional features acquired by GNN are fused together in the proposed network model. The results of the experiments on the synthetic dataset ModelNet40 indicate that the proposed model can obtain a 41.69% reduction in rotational root mean square error and 23.80% reduction in translational root mean square error compared to the state-of-the-art algorithm.