Deep Learning Semantic Segmentation-Based Scan-to-BIM for Indoor Point Clouds

Building information modeling (BIM) is essential for effectively managing reconstructed buildings throughout their operational and maintenance stages, aiding in the progress of smart city projects. However, conventional scan-to-BIM techniques often depend on manual or semi-automated processes and do not fully exploit the rich semantic details of point clouds. As a result, these methods encounter difficulties in accurately modeling large, intricate indoor environments, leading to inefficiencies in BIM creation. To overcome these challenges, this paper presents a novel scan-to-BIM pipeline that integrates deep learning approaches and laser scanner point clouds. The proposed method consists of several key stages, such as preprocessing of scanned data, semantic segmentation through an improved deep learning model, clustering of room density, detection of room clusters and segments, extraction of segment features, identification of doors and windows, and the automated generation of BIM models. These processes work together to automatically create precise and detailed 3D models of building structures. The pipeline is capable of handling large datasets, accommodating both Manhattan and non-Manhattan geometries, and employing a parametric BIM algorithm in Revit software for automated reconstruction. The pipeline has been validated with point cloud data of various indoor layouts, demonstrating outstanding performance in reconstructing indoor elements, with precision and recall values of 97.60% and 96.70%, respectively. The resulting BIM models show excellent completeness and geometric accuracy, surpassing traditional methods in both efficiency and quality.