Efficient and Accurate Method of Point Cloud Registration Based on Plane Correspondences for Structured Scenes

Registration of point clouds is a crucial process in the domain of Light Detection and Ranging (LiDAR). To enhance both efficiency and accuracy in registering point clouds in structured scenes, this paper proposes a registration method based on plane correspondences by enhancing the performance of plane extraction and plane matching without initial guesses from odometry or LiDAR. The proposed registration method defines a novel planar feature descriptor based on the distribution and number of points to improve the plane extraction module, effectively filtering out extraneous planes. Furthermore, a novel angular feature descriptor based on geometric information and the planar feature descriptor, are incorporated to enhance the accuracy of the plane matching, in addition to the traditional feature descriptor. The results obtained from the experiments indicate that the proposed registration method outperforms other existing algorithms, achieving success rates of 98%, 97%, and 100% on the Apartment, ETH, and Stairs datasets, respectively.