LiDAR-based global 3D pose measurement via synthetic scan retrieval and hierarchical registration
TL;DR
This paper proposes a synthetic-scan-based hierarchical LiDAR measurement framework for fast 3D global relocalization that substantially reduces online computation while maintaining high pose-measurement accuracy.
Abstract
Accurate pose measurement without an initial pose estimate is essential for mobile robots operating in GNSS-denied or structurally complex environments. However, global 3D relocalization from LiDAR measurements remains computationally demanding because direct registration over 3D prior maps involves a high-dimensional search space and dense point-cloud matching. To address these issues, this paper proposes a synthetic-scan-based hierarchical LiDAR measurement framework for fast 3D global relocalization. In the offline stage, feasible sensor positions are generated in a 3D occupancy grid using constraint-aware rapidly-exploring random tree-based uniform sampling, and at each sampled position, first-return ray casting is performed to generate a synthetic LiDAR scan whose descriptor is stored together with the corresponding position in the descriptor database. In the online stage, the retrieved 3D position and yaw are combined with nominal roll and pitch values to initialize iterative closest point (ICP)-based point-cloud registration. ICP subsequently refines the rigid transformation in SE(3) and outputs the final pose estimate. Real-world experiments demonstrate that the proposed method achieves pose-wise mean relocalization times of approximately 3 s or less and an average position error of 8 cm in 3D environments. Compared with conventional global registration and feature-matching baselines, the method substantially reduces online computation while maintaining high pose-measurement accuracy.