Semantic Segmentation of LiDAR Point Clouds for 3D Mapping of Underground Space

Abstract

Underground space is among the most challenging environments for 3D mapping because the Global Navigation Satellite Systems (GNSS) signals are often inaccessible. This thesis investigates the use of the LiDAR-based Simultaneous Localization and Mapping (SLAM) technology to map such underground space. Underground parking lots, as an example, offer valuable solutions to the challenges posed by growing populations and urbanization, such as limited surface area, traffic congestion, and environmental concerns. They are GNSS-denied, geometrically repetitive, highly occluded by vehicles and pillars, and contain large, low-texture and specular surfaces that degrade sensing and registration. To support rigorous evaluation under these conditions, this thesis contributes three site-specific underground parking datasets captured using a hand-held LiDAR device, GeoSLAM. Each dataset provides clean point clouds and semantic labels for the core structural and operational classes: wall, pillar, vehicle, and ground, enabling controlled benchmarking. Since low-cost LiDAR scans yield sparse, non-uniform point distributions that omit fine structural features, the first study of the thesis addresses point cloud upsampling, an essential step for creating high-definition maps that preserve fine structural details while ensuring uniform data distribution for downstream tasks. Five deep learning upsampling models including PU-Net, PU-GAN, PU-GCN, PU-Transformer, and RepKPU are trained and tested in a unified pipeline and evaluated with Chamfer Distance for average surface fidelity, Hausdorff Distance for worst-case deviation, and inference time for deployability. RepKPU consistently delivers the best accuracy–latency trade-off in underground setting. Since accurate semantic understanding is crucial for structure-aware mapping and autonomous navigation in complex indoor environments, the second and third studies target semantic segmentation for underground parking spaces, first using Transformer-based backbones and then extending the evaluation to Mamba-based architectures. For Transformer-based methods (PT, PCT, and 3DGTN), the generalization across the three different parking lots is assessed using overall accuracy (OA), mean Intersection over Union (mIoU), and F1-score. The results establish 3DGTN as the most accurate and stable Transformer framework across all three sites. Complementing the Transformer study, Mamba-based methods (PointMamba, PoinTramba, and 3D-UMamba) are compared on the same datasets with 3D-UMamba offering the best overall performance.

point cloud upsampling semantic segmentation GNSS-denied environment underground parking lot Transformer-based methods Mamba-based methods