Multi-threaded GPU-accelerated LiDAR–IMU Localization and Mapping Framework for High-speed Autonomous Driving

High-speed autonomous driving requires a localization framework that is both computationally efficient and robust against sensor degradation and global navigation satellite system (GNSS) denied conditions. This paper presents a multi-threaded graphics processing unit (GPU)-accelerated light detection and ranging (LiDAR) – inertial measurement unit (IMU) localization and mapping framework designed to achieve real-time, high-precision performance under high-speed driving. The proposed system integrates adaptive sliding-window mapping, GPU-accelerated voxelized generalized iterative closest point (VGICP) scan matching, and a score-adaptive extended Kalman filter (EKF) to ensure reliable pose estimation at a 20 Hz LiDAR input and 100 Hz EKF update rates. In the mapping stage, a geo-referenced LiDAR map is generated using a GPU-based scan context pose graph optimization (SC-PGO) pipeline to provide global consistency. The multi-threaded architecture (Worker–Buffer–Callback) enables the parallel execution of LiDAR scan matching, map extraction, and EKF prediction–update cycles, significantly reducing latency compared to conventional single-threaded pipelines. Real-world experiments using the Dallara AV-24 platform at the Laguna Seca Raceway demonstrate that the proposed framework achieves robust localization even in GNSS-denied regions such as tunnels and bridge-covered segments.