Automatic coarse registration of point clouds using plane contour shape descriptor and topological graph voting

Registration is a basic yet crucial task in point cloud processing. Obtaining robust as well as accurate features and correspondences is critical to feature-based point cloud registration. Aiming at structured urban scenes, this paper employs plane as primitive for feature extraction and proposes an automatic coarse registration algorithm of pairwise point clouds based on plane contour shape descriptor and topological graph voting. It contains three major steps: First, in order to improve the efficiency and accuracy of plane extraction, an efficient voxel-based plane segmentation algorithm is applied, then a RANSAC plane fitting method is used to obtain high precision plane parameters. Second, we propose a plane contour shape descriptor, hole inner angle combined with triangular centroid distance (HIA-TCD), for obtaining corresponding planes in two laser scans, which is robust to the contour shape matching in the urban scene. Finally, according to the position relationship between plane correspondences, a topological graph is built, and an edge voting strategy is proposed to eliminate potential unreliable plane correspondences, then the optimal matched planes are selected to estimate the rigid transformation. The algorithm is tested by three sets of large-scale outdoor and two sets of indoor scene data, and the experimental results show that the presented algorithm can effectively register the pairwise point cloud. In the outdoor scene, the rotation errors are less than 0.3 degrees, and the translation errors are less than 0.1 m. In the indoor scene, the rotation errors are less than 0.2 degrees, and the translation errors are less than 0.1 m. In addition, our algorithm is superior to the state-of-the-art in efficiency.