Automated and Semantic-Rich BIM Modeling of MEP Systems Using Monocular Cameras

Accurate and updated as-built building information modeling (BIM) has become an effective tool for managing mechanical, electrical, and plumbing (MEP) systems due to its capabilities in enhancing visualization and data interoperability. Video-based 3D reconstruction methods offer a cost-effective and portable solution for creating as-built BIM models. However, the process of extracting images from video often results in low-quality input data due to issues such as blurriness and distortion, and the low accuracy and density of reconstructed point clouds results in challenges for segmentation and detailed modeling. Using monocular cameras only, this paper presents a novel, automated BIM modeling approach that overcomes these challenges. The newly designed method consists of three compartments: an adaptive frame extraction method that efficiently generates high-quality image sequences from video; the key image identification and semantic transfer techniques that simplify semantic segmentation by converting it into a 2D task on selected images; and the undirected graph-based modeling approach combined with prior knowledge that can produce semantically rich BIM models. The method was tested across multiple scenarios. Results demonstrated its reliability in generating accurate as-built BIM for MEP systems from videos. The method fills an important gap in automated BIM modeling to provide accurate as-built information for managing MEP systems.