Learning MFSC: A Robust and Lightweight Descriptor via Unsupervised Multi-Feature Fusion for LiDAR Loop Closure

Loop closure detection plays a crucial role in correcting drift and ensuring long-term robustness in LiDAR-based SLAM systems. However, existing handcrafted global descriptors such are often sensitive to viewpoint changes, environmental dynamics, and rely on limited feature information. In this work, we propose MFSC, a novel global descriptor generated via an unsupervised learning framework. MFSC enriches the representation by incorporating multiple complementary LiDAR features and utilizes a lightweight learned design. This unique combination achieves crucial rotation invariance while enabling highly efficient similarity computation for fast descriptor matching. Extensive evaluations on the KITTI dataset demonstrate that MFSC significantly outperforms common handcrafted baselines, highlighting its suitability as a practical, high-performance, and data-label-free solution for large-scale LiDAR loop closure detection.