Sequential Semantic Descriptor from 3D Point Clouds for Place Recognition

The closed-loop detection of 3D point clouds remains a significant challenge due to the complexities involved in generating effective descriptors that are robust to occlusion and viewpoint changes. Unlike most existing methods focusing on extracting local, global, and statistical features from the raw point clouds of a single frame, our approach emphasizes the utilization of semantic-level scene graphs derived from successive multiple frames, with superiority in robustness to environmental changes. Drawing inspiration from human cognitive processes that facilitate scene cognition by identifying semantic objects and continuous observation, this study introduces an innovative closed-loop detection method based on sequential semantic graphs. First, we propose a novel semantic graph representation method for point clouds scenarios by preserving the semantic and topological information of the original point clouds. Then, we represent a scene with consecutive multiple frames of semantic graphs. Thus, the closed-loop detection is modeled as a multigraph similarity matching problem. A fast and efficient multigraph similarity network is then designed to calculate the similarity. An exhaustive evaluation of the KITTI dataset shows that our method substantially outperforms state-of-the-art methods, exhibiting robust performance in the face of occlusion and viewpoint variation, with superior performance on challenging reverse closed-loop problems.