A Global Point Cloud Semantic Mapping Method for Underground Spaces Under Extreme Spatial Constraints

Narrow and irregular underground roadways commonly suffer from dim lighting, absence of satellite signals, and complex spatial topologies, making it difficult for traditional methods to achieve rapid and accurate spatial perception and semantic modeling in such environments. Existing studies mostly focus on large-scale underground spaces such as tunnels and subways, where datasets feature single scenes and limited scales and cannot support point cloud understanding tasks under extreme spatial constraints. Thus, insufficient exploration has been conducted on small, narrow, and geologically complex underground roadways. To address this research gap, this study has constructed the NUH-PCD (Narrow Underground High-fidelity Point Cloud Dataset), which is the first point cloud dataset for ultra-narrow and irregular underground roadways. This study employs Simultaneous Localisation and Mapping (SLAM) integrated with Real-Time Kinematic (RTK) technology to achieve high-density point cloud data acquisition and geometric modelling. The core focus lies on point cloud semantic segmentation and the generation of global semantic maps. Through point cloud semantic segmentation techniques, unstructured point clouds are transformed into global maps rich in semantic information, providing technical support for the monitoring of underground space conditions. The PointNet++_DAG semantic segmentation model was constructed by innovatively embedding the dynamic attention gating (DAG) module into the PointNet++ framework, enabling accurate segmentation of eight core categories of targets in underground roadways and further generating global semantic maps with abundant semantic information. Experimental results demonstrate that the proposed PointNet++_DAG model achieves a mean intersection over union (mIoU) of 76.73%, which is 14.51 percentage points higher than that of the baseline PointNet++ model. This work effectively breaks through the bottleneck of global point cloud semantic mapping for underground spaces under extreme spatial constraints, offering an efficient solution for intelligent cognition of underground roadway environments.  It also lays the foundation for subsequent safety monitoring and emergency response, possessing significant theoretical value and engineering application prospects.