Point cloud registration algorithm based on feature extraction and improved ICP

Point cloud registration is an important part of three-dimensional point cloud processing, but as a widely used iterative closest point (ICP) algorithm, it has high requirements on the initial pose and overlap rate of source and target point clouds, and is prone to local optimization. Therefore, a point cloud registration algorithm based on feature extraction and improved ICP was proposed, which mainly included two core parts: sample consensus initial alignment (SAC-IA) and ICP fine registration. Firstly, the incoming point cloud was preprocessed by OCtree accelerated voxel filtering to remove discrete points and streamline the point cloud. Then, the fast point feature histogram (FPFH) feature descriptor was calculated for rough registration of SAC-IA algorithm. Finally, based on the initial matrix obtained by rough registration, the normal vector constraint information was introduced to improve the ICP algorithm for fine registration. The algorithm was verified on the public data set. The results show that the proposed algorithm effectively improves the registration accuracy and efficiency compared with the classical ICP algorithm, and provides a feasible method for the subsequent 3D reconstruction and other aspects.