LCDNet: Deep Loop Closure Detection and Point Cloud Registration for LiDAR SLAM

Loop closure detection is an essential component of simultaneous localization and mapping (SLAM) systems, which reduces the drift accumulated over time. Over years, several deep learning approaches have been proposed to address this task; however, their performance has subpar compared handcrafted techniques, especially while dealing with reverse loops. In article, we introduce novel loop network (LCDNet) that effectively detects closures in light ranging (LiDAR) point clouds by simultaneously identifying previously visited places estimating six degrees freedom relative transformation between current scan map. LCDNet composed a shared encoder, place recognition head extracts global descriptors, pose estimates two clouds. We based on unbalanced optimal transport theory implement differentiable manner allow for end-to-end training. Extensive evaluations multiple real-world autonomous driving datasets show our approach outperforms state-of-the-art cloud registration techniques large margin, Moreover, integrate into LiDAR SLAM library provide complete system demonstrate generalization ability using different sensor setup unseen city.