Localizability-aware Point Cloud Registration for Autonomous Racing

Abstract

LiDAR-based localization for autonomous racing requires fast and reliable pose estimation, especially in challenging racing track environments. Traditional methods like particle filters and Iterative Closest Point (ICP) algorithms often fail when navigating through regularly patterned tracks, such as straight hallways, where the lack of distinct geometric features makes pose estimation difficult. To ensure fast recovery from such failures, this paper introduces a novel scan-to-map point cloud registration method that jointly leverages the localizability of scan data and the visibility of map data. This approach en- hances the accuracy of correspondence matching during the registration process, leading to improved localization performance. Extensive numerical simulations demonstrate that the proposed method sig- nificantly outperforms existing techniques in terms of localization recovery, particularly in challenging environments.