Scan Context++: Structural Place Recognition Robust to Rotation and Lateral Variations in Urban Environments

TLDR

Place recognition is essential for robotic navigation, yet most research focuses on visual appearance‑based methods. This work aims to enable structural place recognition using range sensor data. We extend a rotation‑invariant spatial descriptor to be robust to both heading and translation, add two sub‑descriptors, and combine topological retrieval with 1‑DOF semimetric localization to bridge topological and metric localization. The method was evaluated across varying environmental complexity and scale, and its source code is publicly available for easy integration into LIDAR‑SLAM systems.

Abstract

Place recognition is a key module in robotic navigation. The existing line of studies mostly focuses on visual place recognition to recognize previously visited places solely based on their appearance. In this article, we address structural place recognition by recognizing a place based on structural appearance, namely from range sensors. Extending our previous work on a rotation invariant spatial descriptor, the proposed descriptor completes a generic descriptor robust to both rotation (heading) and translation when roll–pitch motions are not severe. We introduce two subdescriptors and enable topological place retrieval followed by the 1-degree of freedom semimetric localization, thereby bridging the gap between topological place retrieval and metric localization. The proposed method has been evaluated thoroughly in terms of environmental complexity and scale. The source code is available and can easily be integrated into existing light detection and ranging simultaneous localization and mapping.