A Loop Closing Network for Place Recognition and Loop Poses Estimation Tasks in LiDAR SLAM

Loop closing plays a vital role in SLAM for robotic system, as it can be employed to mitigate accumulated drift. However, for 3D LiDAR SLAM, certain current approaches might struggle to accurately identify loops and often cannot rectify the complete 6-DoF loop poses, particularly when en-countering challenging scenarios like reverse loops. To address these limitations, we propose a loop closing network named LC-Net, which is designed to perform both place recognition and loop poses estimation tasks. Within the framework of the proposed method, the feature of keypoints is extracted and fed separately into the VLAD module and the RANSAC module to accomplish respective tasks. The evaluation indicates that the proposed method achieves high efficiency and outperforms some existing approaches. Additionally, we incorporate the proposed method into the LiDAR SLAM system, effectively enhancing trajectory accuracy.