Knowledge-driven inference for automatic reconstruction of indoor detailed as-built BIMs from laser scanning data

Abstract

In spite of recent advances, a significant disparity between automatically reconstructed building information models (BIMs) and real scenes persists, particularly within complex indoor environments. With this objective in mind, a knowledge-driven as-built BIM reconstruction method is proposed, which employs laser scanning point clouds and panorama images. Initiation of the method involves the segmentation of 3D data into individual rooms through the application of wall constraints. Subsequently, geometric regularization of the rooms is performed, accompanied by the establishment of topological relations through the maintenance of rigid consistency among rooms. Finally, building frames and components embedded within walls are reconstructed and assembled to formulate comprehensive BIMs. The method was evaluated using indoor point cloud datasets (ISPRS and WUT), showing that it outperforms state-of-the-art techniques with room segmentation accuracy and completeness of 0.98 and 0.88, respectively, and achieving millimetre level accuracy on average.