Nonrigid Point Set Registration Algorithm Based on Alternating Triangle and Translation Feature Descriptors

Non-rigid point set registration (NRPSR) is a fundamental research area in the domains of robotics and industrial measurement, aiming to estimate the correspondence and spatial transformation between two point sets. In reality, severe deformation, numerous outliers, large-angle rotation, and noise will be generated between two point sets. The feature descriptor (FD) can effectively extract the point set feature under data degradation disturbances, but most existing FDs are mainly directed at one or two disturbances mentioned above. Hence, this paper proposes an NRPSR algorithm based on two novel alternating triangle and translation FDs, which still maintains robustness under most disturbances. The proposed two local FDs are based on pseudo-correspondence, Euclidean distance, translation vector, triangular area and angle to extract the local feature difference between two point sets. The two proposed FDs are then combined through a transition method to construct the alternating triangle and translation local FDs, which can ensure the real-time performance. Then, the local and global feature differences are combined to construct the correspondence cost function, which is solved by the linear assignment algorithm. Finally, the thin plate spline function and simulation annealing algorithm are utilized to solve the spatial transformation and algorithm iteration, respectively. Experimental results demonstrate that the proposed FDs have strong descriptiveness and robustness, and the performance of our algorithm exceeds the state-of-the-art algorithms under most cases of severe data degradation.