SGTD: A Semantic-Guided Triangle Descriptor for One-Shot LiDAR-Based Global Localization

Abstract

This paper presents a novel one-shot global localization algorithm based on semantic-guided triangle descriptors to address initialization and global localization challenges in GNSSdenied environments. By encoding semantic geometric information into triangle descriptors, the proposed approach achieves robust 6-DOF global localization. The algorithm frst employs a Dynamic Curvature Voxel Clustering method to extract stable semantic instances from LiDAR point clouds, effectively mitigating the impact of dynamic objects on localization accuracy. Then, semantic instance objects are selected based on a fully static strategy to construct semantic-guided triangular descriptors. Global localization is achieved through triangle matching via composite binary coding with a pre-built semantic hash map database, followed by a voting mechanism to identify candidate frames. In the geometric verifcation phase, an initial pose estimate is obtained through Singular Value Decomposition, which is subsequently refned through point cloud registration to achieve precise 6-DOF pose estimation. Extensive experiments conducted on multiple datasets, including the campus environments MCD and MulRan, as well as the urban environment Apollo, demonstrate that our approach signifcantly outperforms existing state-of-theart methods in terms of both robustness and accuracy across diverse scenarios. Our method has been open-sourced and can be accessed at: https://github.com/Hfx-J/SGTD.git.