Robust LiDAR-based loop closure detection method for enhanced 3D reconstruction consistency

High-consistency 3D reconstruction from LiDAR is essential for autonomous driving. However, accumulated pose drift during the reconstruction process often leads to global structural inconsistency. This paper presents a robust loop closure detection (LCD) method that integrates a skeleton-projection geometric descriptor with a Siamese network enhanced by evolutionary feature augmentation. An alternating optimization mechanism jointly refines the network and data augmentation strategies, enabling adaptive learning under viewpoint and environmental variations. The proposed method has been extensively validated on both public and self-collected datasets, demonstrating its effectiveness and robustness in large-scale 3D reconstruction.