Centralized RANSAC-Based Point Cloud Registration With Fast Convergence and High Accuracy

Abstract

For point cloud registration, the purpose of this article is to propose a novel centralized random sample consensus (RANSAC) (C-RANSAC) registration with fast convergence and high accuracy. In our algorithm, the novel contributions are, first, the proposal of a scale histogram-based outlier removal to delete outliers from the initial line vector set $\mathbf {L}$ for constructing a reduced line vector set $\mathbf {L}_{\text{red}}$; second, the handshake cooperation between the host RANSAC (H-RANSAC) only working on L and the local RANSAC (LCL-RANSAC) only working on $\mathbf {L}_{\text{red}}$; third, in each handshake process, after receiving the global registration solution and the global iteration number $x_{H}$ from H-RANSAC, LCL-RANSAC uses the received global solution as the initial solution of the modified TEASER++ (M-TEASER++) method to calculate its first local registration solution. If the first local registration solution satisfies the global iteration number inheritance condition, LCL-RANSAC directly sends the accumulated iteration number, $x_{H} + 1$, and the first local solution back to H-RANSAC; otherwise, LCL-RANSAC iteratively refines its local solution using the M-TEASER++ method, and then sends the resultant local solution and the required local iteration number $x_{\text{LCL}}$ to H-RANSAC for updating the global solution, the global iteration number to $x_{H} := x_{H} + x_{\text{LCL}}$, and the global confidence level. Due to $|L_{\text{red}}| \ll |L|$ and employing the global iteration number inheritance condition test into our algorithm, we have conducted extensive experiments on testing point cloud pairs to show the registration accuracy and execution time merits of our algorithm relative to the state-of-the-art methods.