GPU-accelerated embedded EKF-SLAM based on LiDAR–IMU–GNSS fusion for autonomous ground vehicles

Abstract

Abstract Simultaneous Localization and Mapping (SLAM) is a key capability for autonomous navigation, yet it remains computationally demanding, especially when integrating multiple sensors such as LiDAR, IMU, and GNSS, each requiring a dedicated processing pipeline. Although these sensing modalities differ, their underlying computations contain many independent operations that can be parallelized. In this work, we present a parallelized and loosely coupled EKF-based SLAM framework that fuses LiDAR, IMU, and GNSS to jointly estimate vehicle pose and the environmental map. The back-end incorporates a Gauss–Newton scan-matching refinement to enhance pose accuracy and ensure map consistency. To meet real-time constraints, we introduce modular OpenCL GPU kernels that offload the most computationally intensive operations, namely scan-matching cost evaluation and Gauss–Newton optimization, while the remaining EKF updates execute efficiently on the CPU. This heterogeneous design significantly improves computational efficiency while preserving estimation accuracy, enabling real-time SLAM on resource-constrained robotic platforms.