Hierarchical and Validated Branch-and-Bound Method for Global Point Cloud Registration

Meng Xu, Jikai Wang, Kezhi Wang, Zonghai Chen

Abstract: Global registration of LiDAR point clouds is a pivotal requirement for autonomous platforms localization, loop closure detection, and map fusion. Despite numerous global registration methods proposed in recent years, their performance is constrained by two primary issues: Low overlap, and large-scale point clouds. These issues become exacerbated when registering partial maps with significant pose discrepancies. To address these issues, we introduce a Branch-and-Bound (BnB)-based method incorporating hierarchical subproblem solution and consistency-based validation, which we term HV-BnB. To mitigate the high time complexity associated with large-scale point cloud registration, our approach reduces the BnB search space by leveraging the Atlanta world assumption and decomposes pose estimation problem into translation estimation and transformation optimization. Hierarchical strategies are also designed to balance between representativeness and accuracy for subproblem solution, thereby enhancing both efficiency and precision. Although BnB algorithm is a full search algorithm, it cannot guarantee global optimal registration under high outlier rate, we propose validation strategy to achieve robust global registration by evaluation the consistency of correspondences. Through verification in both Scan2Scan and Map2Map registration tasks under various scenarios and datasets, our proposed method demonstrates robust and fast global registration performance compared to the state-of-the-art baselines.