LiDAR localization at 100 FPS: A map-aided and template descriptor-based global method

Abstract

With the development of multi-beam Light Detection and Ranging (LiDAR) sensors, fast and accurate LiDAR-based localization has become a crucial issue in robotics and autonomous driving. However, balancing accuracy and efficiency remains challenging in existing methods. In this paper, we propose a super-fast LiDAR global localization approach that can achieve state-of-the-art (SOTA) accuracy with superior efficiency. Our method leverages template descriptors to capture structural environments and approximates the vehicle’s position via map candidate points. Additionally, we create an offline map database to evenly simulate vehicle orientations. We design a loss function to improve localization accuracy. We extensively evaluated the proposed method in public KITTI outdoor sequences and self-collected indoor datasets. The experimental results show that our approach can run at close to 100 frames per second (FPS) on a single-thread CPU, which is much faster than current SOTA methods. Our average absolute translation errors (ATEs) are 0.20m (indoor) and 0.44m (outdoor), and the average localization success rates are 93% (indoor) and 90% (outdoor). The average localization success rates can exceed 97% in large outdoor scenarios with fine-tuned parameters. The source code will be available in https://github.com/ShiPC-AI.

Keywords

• SLAM
• Global localization
• LiDAR
• Template descriptor
• KD tree