Deep Learning–Based Automation of Scan-to-BIM with Modeling Objects from Occluded Point Clouds

Abstract

As-built building information modeling (BIM) currently is regarded as a tool with the potential to manage buildings efficiently in the operation and maintenance phases. However, as-built BIM modeling is a labor-intensive process that requires considerable cost and time in modeling existing buildings. Although active research on scan-to-BIM automation has addressed this issue, previous studies modeled only major objects such as walls, floors, and ceilings, consequently requiring modeling other objects in indoor spaces. In addition, there was a limitation in modeling objects located in the occluded areas of scanned point clouds. Therefore, this study extracted various indoor objects from a point cloud based on deep-learning, and compensated for incomplete object information from occluded point clouds for automating the process of scan-to-BIM. The number of object classes extracted from the semantic segmentation of a deep learning network was increased to 13, and spatial relationships between objects were defined to improve the accuracy of bounding boxes extracted from point clouds. Furthermore, a parametric algorithm was developed to match the bounding boxes and objects in a BIM library to generate BIM models automatically. In a case study involving an office room, the accuracy of the bounding boxes of some object classes improved by as much as 53.33%. The study verified the feasibility of the proposed method of scan-to-BIM automation for the three-dimensional (3D) reality capture of existing buildings.