A High-precision Point Cloud Registration Method Based on Graph Search Point-to-tangent ICP

To address the issues of defects in correspondence selection model and initial value sensitivity in existing ICP (iterative closest point) based point cloud registration methods, a point-to-tangent ICP registration method based on graph search is proposed. Firstly, a graph-based search method is designed to obtain the closest points and a more observation-conformant hypothesis of correspondence point-to-tangent models is proposed. The intersections of simulated laser rays and the tangents of environmental sampling points are used as correspondence, then error formula is constructed and solved. Secondly, a coarse registration method based on pose particle probability is implemented to calculate possible pose transformation particles and score them, and the high-scored transformation poses are used as the initial input pose for fine registration. Finally, the feasibility of the method is verified through experiments and compared with other methods. The results show that, compared to other methods in the experiment, the proposed method can complete the matching of two frames of 2D LiDAR data in a short time (100 ms) and obtain high-precision estimation results.