A LiDAR-based single-shot global localization solution using a cross-section shape context descriptor

Fusing LiDAR and high definition (HD) maps is a feasible way to achieve global localization in GNSS-denied environments, which is necessary in driverless vehicle and robotic technologies. This paper proposes a single-shot global localization solution that uses only a single current scan of a rotating multiple-beam LiDAR sensor to locate its own location and pose with respect to an HD map in the form of georeferenced point clouds. This single-shot global localization solution estimates the state of the current moment without the previous moment state and thus avoids the nonconvergence problems that plague filter-based methods. The proposed solution allows HD maps from diverse LiDAR sensors to be used for global localization and is more robust than existing methods. The proposed solution consists of two procedures: offline preprocessing and online global localization. In the offline procedure, diverse HD maps are preprocessed to construct a global prior map for the localization process. The online global localization procedure includes two elements: place recognition, location and pose estimation. A novel Cross-Section Shape Context (CSSC) descriptor that is highly descriptive and rotation-invariant is proposed for subsequent processes. Two strategies, two-stage similarity estimation and Nearest Cluster Distance Ratio (NCDR), based on the CSSC descriptor are proposed to improve place recognition precision. A Selective Generalized Iterative Closest Point (SGICP) algorithm is proposed to calculate location and pose accurately using the CSSC descriptor. Comprehensive experiments were performed to evaluate this solution. A comparison of the precision-recall curve of multiple scenes, particularly under changed viewpoint scenes, shows that the CSSC descriptor is more robust than existing descriptors. Experimental analysis also confirms that the proposed strategies, two-stage similarity estimation and NCDR, improve place recognition precision. Also, compared to the generalized iterative closest point algorithm, the SGICP algorithm achieved better accuracy by 31% and efficiency by 60%. The proposed solution achieves a mean relative translation error (RTE) improvement of 27% over the OneShot algorithm on the KITTI dataset. The proposed solution had an average 77% improvement over 1σ RTE relative to the benchmark in tests with the long-term localization NCLT dataset. The mean RTE of the proposed solution was 0.13 m using HD maps from different LiDAR sensors. Our code is available at: https://github.com/Dongxu05/CSSC.