Vision-assisted BIM reconstruction from 3D LiDAR point clouds for MEP scenes

Abstract

Mechanical, electrical and plumbing (MEP) system provides various services and creates comfortable environments to residents in cities. To enhance the management efficiency of the highly complex MEP system, as-built building information modeling (BIM) is increasingly adopted in the world. Currently, as-built BIMs are mostly drawn manually by modelers in BIM modeling software referring to point clouds or on-site photos, which is time consuming and labor intensive. This study presents a novel fused BIM reconstruction approach for MEP scenes. The proposed approach makes the best of the rich semantic information provided by images and accurate geometry information provided by 3D LiDAR point clouds. Firstly, a state-of-the-art deep learning model focusing on semantic segmentation is fine-tuned for the MEP dataset, and then RGB images collected with depth camera are segmented with the well-trained model. Secondly, taking the segmented images and the corresponding depth images as input, a semantic-rich 3D map is generated. Thirdly, an instance-aware component extraction algorithm in LiDAR point clouds given approximate object distribution in 3D space is developed. In the component extraction algorithm, a label transfer technique is proposed to firstly determine the rough locations of targeting objects in LiDAR point clouds. Then, accurate component locations are determined for three types of components including irregular shaped components, regular shaped components, and secondary components attached to walls. Finally, the BIM model is reconstructed based on component extraction results. To validate the proposed technique, experiments were conducted in four MEP rooms in a water treatment plant in Hong Kong. It is demonstrated that the proposed technique is more accurate and more efficient with wider range of applications compared to previous BIM reconstruction methods.