Robust Approach for the Registration of Multi-View Point Sets Based on Weighted Motion Averaging

This study proposes a method for improving the registration performance of multi-view point sets based on weighted motion averaging. First，the relative motions of scans is estimated utilizing a pair-wise registration algorithm. Subsequently，a novel method for hierarchically estimating the initial global motions using the cycle consistency of triplets was described. This method calculates the consistency of all triplets in the view-graph of motions and gradually selects the reliable triplets for the initialization and expansion of nodes，avoiding the failures of motion averaging caused by improper initialization. With access to the accurate initial global motions，it filters outlier edges by evaluating the reliability of pair-wise registration results. Therefore，the final relative motions，as the input to the algorithm for weighted motion averaging，is the inlier subset of the initial relative motions. A comparative experiment was performed on the Stanford 3D scanning repository. Using the pair-wise registration algorithm，we obtain 12 relative motions for four range scans with overlap percentage thresholds of 30%，25%，and 20%. Among these，the error in relative motions increases when the overlap percentage threshold is decreased. The experimental results on these relative motions demonstrate that the proposed method improves the performance and robustness of the weighted motion averaging algorithm for multi-view registration. Experimental results reveal that the proposed method can obtain accurate registration results even when the set of the relative motions contains a high number of outliers，which verified the robustness of our method. © 2023 Tianjin University. All rights reserved.