Deep Learning-based 3D Point Cloud Semantic Segmentation for Mechanical, Electrical, and Plumbing Systems in Open Ceiling Building Structure

Deep learning-based semantic segmentation of point clouds is a vital tool for efficiently generating accurate as-built Building Information Models (BIM) by automating the identification and classification of building components. While traditional deep learning models have primarily focused on segmenting visible structural elements such as columns, walls, doors, windows, and furniture, the growing popularity of open ceiling designs presents unique challenges. Open ceiling environments expose Mechanical, Electrical, and Plumbing (MEP) components, including pipes, ducts, and electrical conduits, which are often characterized by overlapping geometries, visual similarities, and intricate spatial arrangements. These complexities raise questions about whether existing deep learning-based semantic segmentation models can effectively classify MEP components in such challenging environments. This study addresses this gap by evaluating the performance of RandLA-Net, a lightweight deep learning-based semantic segmentation model, for accurately classifying MEP components in open ceiling environments. Using LiDAR technology, 3D point cloud data was collected within a building featuring an open ceiling design, where components were manually annotated to train and test the model. The performance of the model was evaluated through metrics such as overall accuracy (OA) and mean Intersection over Union (mIoU). The results showed an OA of 90.86% and a mean IoU of 83.1%, indicating that the existing RandLA-Net model is capable of effectively segmenting MEP components alongside structural elements, even within the challenging context of open ceiling designs. The findings suggest that RandLA-Net has strong potential for addressing the complexities of MEP component classification in open ceiling environments. Furthermore, the study highlights the opportunities for enhancing semantic segmentation methods to achieve even greater precision in as-built BIM generation, supporting improved maintenance planning and operational efficiency in complex architectural settings.