Improved building facade segmentation through digital twin-enabled RandLA-Net with empirical intensity correction model

Abstract

The Architectural Engineering and Construction (AEC) industry can benefit from accurate building facade segmentation, which can provide valuable insights into building maintenance, urban planning, and security efforts. Light Detection and Ranging (LiDAR) sensors are effective in recognizing building facade components from (three-dimensional) 3D point clouds by registering 3D spatial coordinates and radiometric information, which reveal the spectral property of a scanned surface. Although the radiometric information (i.e., intensity feature) can be used for segmentation, its accuracy may be reduced by factors such as scanning geometry and external factors that affect the object’s radiometric information. To address this issue, this study proposes a robust and automated method for segmenting building facade components using LiDAR point cloud data and an empirical-based intensity correction model to ensure proper segmentation. The proposed method employs RandLA-Net, a deep learning model capable of effectively processing large-scale point cloud data, to classify building facade components based on their spatial features combined with corrected intensity features. By incorporating the proposed method into a digital twin, it is possible to perform accurate building facade segmentation and generate valuable insights into the building’s physical condition, energy efficiency, and aesthetic value in real-time. The effectiveness of the proposed method was experimentally validated using a school building facade, which demonstrated significant improvements in the recognition of facade components and highlighted the potential of digital twin-enabled building facade segmentation for the AEC industry.