3D Point Cloud Registration Method Based on Structural Matching of Feature Points

Kezhi Wang, Jianyu Li, Zonghai Chen, Jikai Wang

Abstract: 3D point cloud is an important expression of three-dimensional environment information. The registration of point cloud is the basis of realizing the functions such as localization, map construction and target detection based on 3D point cloud. For better registration, a method based on feature points and their spatial structure properties is proposed, which is divided into two stages: coarse registration and fine registration. Firstly, a feature point extraction network based on PointNet++ and Probabilistic Chamfer Loss is designed to extract robust and highly repetitive feature points. On the extracted feature points, the Super 4PCS method based on the spatial structure features of point cloud is used for global coarse registration. Then, taking the result obtained by coarse registration as the initial solution, the fine registration method based on NDT is used to register the point cloud more accurately. Finally, accurate registration results are obtained. This method combines the advantages of deep learning method and traditional method based on structure of point cloud, and makes full use of the spatial distribution properties of feature points. Experiments show that this method can achieve good registration results in a variety of application scenarios. And the method also shows the application potential of global location for initialization purpose in small-scale environments.