Geometric-Feature-Enhanced Deep Learning for Bridge Point Cloud Segmentation under LiDAR Occlusion

The integration of LiDAR scanning and computer vision has expanded the use of point cloud data across diverse domains, including robotics, autonomous vehicles, and the construction industry. In particular, Scan-to-building-information-modeling (BIM) technologies that generate digital three-dimensional (3D) models from point clouds have become essential for smart construction, especially in cases where original design drawings of aging bridges are missing or manually documented. While automated Scan-to-BIM techniques have advanced significantly in architectural applications, their application to bridges remains challenging due to the complexity, size, and geometric variability of structural components. In prior work, deep-learning-based part segmentation models either excluded bearings from the part segmentation process or, when included, yielded very low part segmentation accuracy for bearings—typically reporting intersection over union (IoU) scores below 10%—due to severe occlusion. To address this issue, we propose a deep-learning-based part segmentation method that incorporates high-dimensional geometric features—anisotropy and the third eigenvalue—to improve part segmentation accuracy of bearings. The proposed model was trained on a combination of simulated and real-world scanned data, and tested on four separate real-world bridge data sets, including one field-scanned and three publicly available scans. As a result, the method improved the IoU for bearings from previously reported values below 10% to up to 53.48%, demonstrating a significant performance gain. These findings highlight the practical applicability of the proposed method in enhancing the reliability and accuracy of automated Scan-to-BIM processes for complex bridge structures.