Tensor Voting Based 3-D Point Cloud Processing for Downsampling and Registration

Abstract

Point cloud registration is related with many significant and compelling 3D perception problems including simultaneous localization and mapping (SLAM), 3D object reconstruction, dense 3D environment generation, pose estimation, and object tracking. A point cloud can be defined as a data format that consists of a combination of multiple points used to identify an object or environment. The aim of this study is to propose a point cloud registration method, which ensures that the point clouds obtained with 3D LiDAR are sampled while preserving their geometric features and the point clouds are registered with high success rate. For this process, it is inspired from the method known in the literature as Tensor Voting, which is originally used to extract geometric features in N-dimensional space. In point cloud registration process, a coarse registration step has been proposed, which focusses on feature registration instead of point registration.