Real-Time and Robust 3D Place Recognition With Adaptive Data Reorganization and Geodesic-Constrained Plane Suppression

The place recognition capabilities for long-term autonomous navigation of mobile robots face multiple challenges, including degradation environments, dynamic disturbances, and limited computational resources of edge devices. This letter addresses the challenge of improving feature extraction efficiency and recognition accuracy in LiDAR-based place recognition by proposing a novel multi-thread pipeline for real-time and robust 3D place recognition with optimized CPU parallel processing, which can reduce computational time significantly. Furthermore, an adaptive data reorganization strategy is presented to improve efficiency, which achieves fast projection data reorganization through pre-marking of input voxels and parallel processing. To enhance geometric matching robustness in degenerate scenes, a plane suppression module is proposed and implemented by integrating geodesic distance and kernel regularization through normal vector manifold embedding and adaptive noise weighting. Experimental validation shows that our method achieves superior accuracy and efficiency compared with STD, ScanContext++, BoW3D in KITTI, NCLT, Wild-Places, and Complex Urban Datasets. Specifically, while maintaining 100% precision, our method achieves a 350% increase in the minimum recall rate and an 11.8% increase in the average recall rate compared to the best results of other methods. Additionally, our method reduces processing time by 54.8% relative to the fastest method (STD) across all sequences. Furthermore, in a real-world coal bunker scenario with a self-built drone, our method demonstrates robustness with a 95.49% successful relocalization rate, calculated from 511 valid frames.