Neural Network Algorithm to 3D Point Cloud Registration Using Virtual Points

Abstract

Three-dimensional point cloud registration is of great importance in robotics and computer vision because it allows finding a geometric transformation to align a pair of point clouds with unknown correspondences. In recent years, deep learning methods have become a major player in the field of computer vision. An essential element of point cloud registration is the correspondence estimation between point clouds. The basic idea is to establish correspondence using multidimensional descriptors for each point. In this paper, we describe a neural network algorithm for registering incongruent point clouds. The proposed algorithm uses virtual points and fuzzy correspondence between clouds, and it is partially based on the PointNet++ neural network. Computer simulation results illustrate the effectiveness of the proposed method.