Large-scale underground mine positioning and mapping with LiDAR-based semantic intersection detection

Various coal mine robots (CMRs) and unmanned aerial vehicles (UAVs) are implemented to explore unknown mines for improving the safety and efficiency of mining. It is challenging for CMRs and UAVs to achieve accurate positioning due to the absence of the Global Positioning System (GPS), poor lighting conditions, and similar geometric features in complex mine scenes. LiDAR-based localization and mapping methods are more accurate than others, while long-time running in large-scale scenarios will introduce nonnegligible cumulative errors. This study presents a semantic-aided Li- DAR simultaneous localization and mapping (SLAM) with loop closure, which leverages the uniqueness of mine intersection structure to establish stable semantic loop closure. Specifically, we propose a semantic intersection descriptor of translation and rotation invariance, which encodes 3D point clouds of the same intersection from different positions and viewpoints into a unified image. By using the semantic descriptor, we can construct a constant loop constraint when the same intersection is revisited from different directions to reduce cumulative drift. We provide experimental validation using large data sets collected in two large underground mines, namely, a simulated Edgar mine deployed in ROS Gazebo and a public underground mine data set provided by ETH. Experimental results show that the proposed method has higher localization accuracy and outperforms the existing LiDAR-based SLAM strategies.