HTD: A Hybrid Triangle Descriptor Combining Geometry and Intensity Information for 3D Loop Detection

Loop detection is essential for a simultaneous localization and mapping (SLAM) system. For LiDAR SLAM, most of existing methods only extract geometrically-based features to search loops, which may lead to false positives when dealing with geometrically similar scenes in different locations, such as corridors with similar layouts on different floors in indoor environments. In this work, we propose a novel global descriptor, hybrid triangle descriptor (HTD), for 3D loop detection. HTD extracts corner keypoints and plane keypoints based on the structural information and the distribution of point intensity on the plane, and then builds geometry triangle descriptors and intensity triangle descriptors based on these keypoints. The triangle descriptors are invariant to translation and rotation, and can provide precise 6- DoF transformation between two loop frames. Besides, we propose an adaptive voxelization strategy to better extract keypoints and employ a hybrid scoring method to enhance the robustness of HTD. In our experiments, we compare our method against other state-of-the-art methods on public outdoor and indoor datasets (i.e., KITTI and TIERS) and our self-collected indoor dataset collected by a solid-state LiDAR. The results show that our method outperforms other state-of-the-art methods on most scenarios, especially on indoor scenarios.