Semantic scan context: Global semantic descriptor for LiDAR-based place recognition

Place recognition plays an important role in a typical simultaneously localization and mapping(SLAM) framework, which allows the autonomous mobile robot to identify the revisited places. Recently, emerging researches focus on incorporating geometry features to achieve accurate and view-invariant relocalization in outdoor environment. However, the ambiguity of geometry features occurs in the scenes with similar objects. To address this problem, we propose Semantic Scan Context, a noval global descriptor based on 3D LiDAR scans and static semantic information such as trunks, poles, traffic signs, buildings, roads and sidewalks. This descriptor not only records the geometrical structure of a 3D LiDAR scan, but also encodes the semantic distribution information. Furthermore, we introduce a coarse-to-fine hierarchical retrieval method to realize the efficient matching for the proposed descriptors. We define three distances to measure the similarity of two places. By weighted sum of these three distances, revisited places can be exactly determined. The recall-precision curve of proposed method is evaluated on public datasets and compared with existing methods. The experimental results proof that our method achieves a competitive re-identification performance.