Semantic Topology and Global Descriptors Based Localization for UGV in Prior Air-View Map

The localization for Unmanned Ground Vehicle (U G V) in prior aerial view map remains a challenging task due to the inherent difficulty in generating descriptors that are both effective and robust to occlusion and viewpoint changes. In contrast to the majority of existing methods that primarily concentrate on extracting local and global features from raw point clouds, our approach combines semantic-level topological relations with global descriptors, thereby providing superior robustness against air-ground viewpoint changes. In this paper, we propose a localization method that is founded on semantic topological relations and global descriptors. First, semantic information, global descriptors and instance information were extracted on the point cloud. A semantic map of the current scan was generated. Second, multi-hypothesis tracking was used to localize ground robots position in the prior aerial view map. We evaluate our approach on the stpls3d datasets. The experimental results show that our method can effectively localization under the air-around viewpoint changes.