Learning-assisted scan-vs-BIM registration and quantification for robust progress monitoring

Automated monitoring of construction progress remains hindered by the difficulty of aligning noisy, incomplete site scans with idealized building information modelling models. Existing approaches often fail under clutter, occlusion and partial assemblies, limiting their reliability in practice. This article presents a robust framework that combines semantic segmentation, iterative registration with geometric filtering and hierarchical, weighted progress quantification. Unlike pipelines that treat segmentation or registration in isolation, our approach leverages semantic structuring to guide the alignment process. It then translates voxel-level overlaps into interpretable, element-wise completion metrics that directly reflect construction progress. Validation on the ISPRS benchmark shows consistently higher correspondence retention than both classical [iterative closest point (ICP), RANSAC + ICP] and deep learning-based (DCP, RPMNet, PointNetLK) methods, achieving top fitness scores in most models. A real-world case study further confirms resilience to noise and clutter, producing a calculated global progress of 93.5% against a real value of 98.39%, with element-wise deviations bounded by a 9.90% standard deviation. These results confirm the framework’s readiness for deployment in digital construction management.