Abstract: Occlusion, noise, and non-orthogonal structures hinder the automated generation of standardized building information modeling (BIM) from indoor point clouds. An end-to-end Scan-to-BIM method integrating semantic priors and global optimization is proposed. Specifically, point-wise semantic information is ob-tained using the Swin3D network, and spatial boundary priors are extracted from ceiling projections. On a 2D plane, multi-stage wall-line extraction and orientation-aware fusion are performed, and a seman-tic-geometric joint probability model is introduced to generate high-confidence candidate wall lines. The room-closure problem is then formulated as a mixed-integer linear programming (MILP) problem to select a topologically optimal wall-line subset under connectivity, loop-closure, and model-complexity con-straints. Based on the optimization results, wall parameters are estimated, and parametric modeling is im-plemented using an open-source modeling tool compliant with the Industry Foundation Classes (IFC) standard to output IFC-compliant BIM models. Experiments on the BIMNET dataset demonstrate that, compared with Cloud2BIM and A-Scan2BIM, the proposed method achieves lower Chamfer and Hausdorff distance errors, with a projected intersection-over-union (IoU) of 0.758 and wall-segment precision of 0.799. At the topological level, the room-closure rate increases to 0.71 and the corner error decreases to 0.037 4 m. Ablation experiments further verify that semantic constraints, ceiling priors, and MILP-based global optimization contribute to the method's geometric accuracy and topological completeness. The pro-posed method provides a reproducible and standards-compatible technical pathway for automated Scan-to-BIM in complex indoor scenes.