Real-Time and Robust LiDAR-Centric Localization System for High-Speed Autonomous Racing

Abstract

Autonomous racing imposes stringent requirements on localization because vehicles operate at high speeds on large-scale tracks under tight safety constraints. These operating conditions make general-purpose LiDAR odometry and simultaneous localization and mapping (SLAM) difficult to apply directly. Incremental estimation can accumulate drift and degrade global consistency. Sensor-stream dropout and temporal inconsistency may interrupt localization continuity, while dense multi-LiDAR scans increase pose-update latency. To address these challenges, this work presents a LiDAR-centric real-time relocalization framework for known racing tracks. It adopts a prealigned global map and hash-based local map retrieval for continuous pose correction, eliminating the need for online mapping and loop closure. A global navigation satellite system (GNSS)-disciplined unified time base aligns LiDAR-inertial measurements, while redundancy-aware sensor selection maintains fixed-dimensional estimator inputs during partial sensor-stream dropout. Furthermore, graphics processing unit (GPU)-accelerated point cloud deskewing and constrained preprocessing mitigate high-speed motion distortion and reduce the registration overhead of dense scans. Real-vehicle experiments were conducted at the Yas Marina Circuit at speeds up to 70 m/s. In laboratory replay of the recorded real-vehicle data, the system achieved a mean GNSS-relative trajectory deviation of 0.765 m and an average registration time below 20 ms for scans containing approximately 150 000 points. It also maintained localization continuity in the tested sensor-stream dropout cases when sufficient constraints remained.