Hierarchical Primitive and Semantics aided Scan Context for place recognition using LiDAR and Monocular Image

Indoor localization involves a challenging and essential task of recognizing places, which has been approached through multi-sensor solutions. However, methods based on a single level of features and homogenous features suffer from ambiguity and lack robustness to changes in environments and viewpoint. To address this challenge, we propose hierarchical primitive, and semantics aided scan context that uses a hierarchical feature comprising primitives and point-level features based on a coupled LiDAR and visual camera system. The proposed feature provides a combination of local and global description, incorporating their advantages while balancing their individual drawbacks. Planar primitives from both image and point clouds are detected for coarse recognition and selection of similar candidates, improving the independence of viewpoint and scenario similarity. Point descriptors, including scan context and SIFT, are then obtained for stage of fine recognition within the previous candidates. The results are evaluated using one real-world indoor dataset. Experimental results demonstrate that the proposed feature descriptor achieves accurate place recognition at the state of the art level, compared to the original scan context descriptor.