Semantic instance segmentation and automated 3D BIM reconstruction for viaduct using LiDAR point clouds and weakly-supervised learning

Abstract

3D reconstruction of Building Information Models (BIM) for transport infrastructure is challenging due to point cloud incompleteness, uneven density, and variations in structural configurations. This paper presents an AI-based semantic instance segmentation approach that leverages weakly-supervised learning for high-precision segmentation and automated BIM reconstruction of transport infrastructure, focusing on viaducts. The method integrates semantic instance segmentation with voxel-based downsampling and density-based filtering to mitigate data incompleteness and uneven density. Mathematical formulations and algorithms are presented, combining geometric representations and spatial relationships of viaduct components to support BIM modelling. A key contribution consists of integrating weakly-supervised learning to segment uneven, incomplete and structurally diverse point clouds, followed by mathematically grounded formulations for high-precision 3D modelling. Experiments demonstrate that the proposed method achieves 94.72 % overall accuracy and 90.51 % mIoU for segmentation, and BIM accuracy exceeding 85 % within 10 mm tolerance between point clouds and generated models, improving BIM reconstruction of transport infrastructure.