Robot localization is a foundational technology for autonomous navigation, enabling task execution and adaptation to dynamic environments. However, failure to return to the correct pose after power loss or sudden displacement (the “kidnapping” problem) can lead to critical system failures. Existing methods often suffer from slow relocalization, high computational cost, and poor robustness to dynamic obstacles. We propose a novel inertial measurement unit (IMU)-LiDAR fusion relocalization framework based on Gaussian historical constraints and adaptive likelihood field matching. By incorporating IMU-derived yaw constraints and modeling historical poses within a 3σ Gaussian region, our method effectively narrows the LiDAR search space. Curvature and normal vector-based feature extraction reduces point cloud volume by 50–70%, while dynamic obstacle filtering via multi-frame differencing and neighborhood validation enhances robustness. An adaptive spiral search strategy further refines pose estimation. Compared to ORB-SLAM3 and adaptive Monte Carlo localization (AMCL), our method maintains comparable accuracy while significantly reducing relocalization time and CPU usage. Experimental results show a relocalization success rate of 84%, average time of 1.68 seconds, and CPU usage of 38.4%, demonstrating high efficiency and robustness in dynamic environments.
Shen et al. (Mon,) studied this question.