SE-VPR: Semantic Enhanced VPR Approach for Visual Localization

Visual Place Recognition (VPR) is a significant task for robotics and autonomous system. However, existing VPR methods ignored the high-level semantic scene information and performed poorly in the symmetrical scene. This paper introduces SE-VPR, which encodes the relative geometric relationship among semantic regions into image global descriptor to overcome the confusion in the symmetrical scene and improve the recall of the coarse localization step of the hierarchical localization paradigm. The proposed VPR algorithm is evaluated and analyzed on the real dataset to verify its effectiveness and superiority. Finally, the dataset used in the experiment is disclosed, which can be used to evaluate the localization performance of the VPR algorithm and hierarchical localization algorithm in complex scenes such as symmetrical and repetitive structures.