Comparative study of point cloud registration approaches

Abstract

Point cloud registration is to obtain a rigid transformation between two different point clouds collected by radar sensors or depth cameras. As a fundamental step in many processes such as reconstruction or segmentation tasks, and Simultaneous Localization And Mapping (SLAM). However, due to the point clouds used in different proposed methods are always collected privately, and the algorithms with different mechanisms will not be compared, there are seldom articles comparing the registration methods systematically.

The purpose of this paper is to compare point cloud registration methods with different mechanisms which mainly include local registration, global registration, and learning-based registration. These methods will be tested on a well-defined combined point cloud benchmark and two classic public point cloud datasets and results will contain multi-level metrics.

In addition, to simulate a more complete actual use case, I also built a virtual SLAM process on gazebo, and obtained point clouds of the scene by A-loam. The virtual environment will contain an indoor scene and an outdoor scene. Different point cloud registration will be used to match two point clouds obtained by two robots in the same scene and the results will be visualized for a more intuitive presentation.