The 3D Lidar Point Cloud Descriptor Based on Structural Unit Soft-Encoding in Structured Environment

Due to the difficulty in descriptor calculation, loop closure detection using 3D lidar point cloud data is a challenging problem in simultaneous localization and mapping (SLAM) systems. Therefore, a novel 3D lidar point cloud descriptor based on structural unit soft-encoding is proposed in this paper, which can be applied to loop closure detection in structured environments. For the difficulty in extracting 3D line segments caused by the sparsity and independence of 3D lidar point cloud, the line segments perpendicular to the ground are extracted firstly by geometric filtering to preserve the structural information of 3D space. Then, a set of structural units is constructed based on the space geometric relationship among the obtained line segments, and the feature vectors are calculated by soft encoding, which are used as the descriptors of the 3D lidar point cloud, Finally, loop closure detection is achieved through the descriptors matching of the two frames of point cloud. Comparative experiments on the KITTI dataset and self-collected dataset show that the proposed approach presents superior performances over the state-of-the-art 3D lidar loop closure detection methods in terms of effectiveness and robustness.