LiDAR-Based Semantic Place Recognition in Dynamic Urban Environments

LiDAR promotes the application of simultaneous localization and mapping (SLAM) in intelligent vehicles. Place recognition is an important component of SLAM. In dynamic urban environments, most existing methods still suffer from changes in sensor viewpoints and scene appearances. To this end, inspired by the perception of humans, who recognize a revisited scene through constant topological relationships of static objects in scenes, we propose a LiDAR-based semantic method for place recognition. First, we design a static object segmentation module (SOSM) with the aid of boundary detection. The module comprises multiple feature fusion enhancement blocks (FFEB) that fuse semantic and boundary features at different scales to optimize the segmentation result. Then, we design a topological graph generation module based on the geometric information and spatial relationships of static objects. The module develops an enhanced topological graph to represent a scene and transforms the place recognition problem into a graph matching problem. Finally, we design a graph similarity estimation module (GSEM) for place recognition. The module extracts high-dimensional global features from the graph. Extensive evaluations on KITTI and real vehicle datasets demonstrate that our method is impressively robust to changes in sensor viewpoints and scene appearances and accurately recognizes similar scenes in dynamic urban environments.