Overview of deep learning-based LiDAR SLAM

LiDAR Simultaneous Localization and Mapping (SLAM) has been successfully applied in fields such as computer vision, autonomous driving, and robotics. Integrating deep learning into LiDAR SLAM can improve algorithmic performance and enhance robots’ autonomous scene understanding, making it an active research topic. This paper reviews four core functions of LiDAR SLAM systems: feature extraction, point cloud registration, loop closure detection, and semantic map construction. Representative deep learning algorithms used in these modules are analyzed and compared. For feature extraction, methods are categorized according to point cloud processing strategies, and key optimization directions at different processing stages are summarized. End-to-end frameworks for point cloud registration are then reviewed. Deep learning-based loop closure detection methods using raw point clouds, projection, and multi-view representation fusion are introduced. Semantic map construction methods and their limitations are also discussed. Finally, typical applications of LiDAR SLAM integrated with deep learning are presented, and future research directions are suggested.