Loop Closure Detection Based on Local and Global Descriptors with Sinkhorn Algorithm

This paper presents a novel loop closure detection pipeline for SLAM systems, addressing the limitations of current deep learning methods in maintaining 3D point cloud structure and extracting high-quality semantic features. We utilize U-Net and FPN for feature extraction, with a descriptor generator that learns from local descriptors. The Sinkhorn algorithm is incorporated for 6DOF transformation matching between point clouds, effectively managing occlusions and aligning source and target clouds. Our method, evaluated on the KITTI dataset, outperforms traditional and other deep learning methods in computational efficiency and real-time performance.