Integrating hierarchical segmentation and vision-language reasoning for spatially complex and occluded MEP point clouds

Abstract

3D BIM reconstruction for MEP systems reduces manual documentation and enhances asset information management. However, the complexity of real-world MEP scenes, characterized by their non-linear trajectories within dense and cluttered environment, frequent data incompleteness due to occlusions, poses significant challenge for instance segmentation and geometric modeling. This paper proposes a hierarchical and progressive segmentation framework that integrates deep learning-based semantic segmentation, geometry-driven instance segmentation, and vision-language model-assisted (VLM-assisted) segmentation refinement. The semantic segmentation module achieves an overall accuracy of 87.02 % and mIoU of 69.10 %, with true positive rates exceeding 97 % for pipe, duct, and tray systems. A voxel-based DBSCAN algorithm is developed to enhance clustering stability and efficiency, followed by an improved RANSAC to extract directional primitives. In addition, VLM-assisted 2D projection analysis is introduced to refine segmentation boundaries and support downstream geometric modeling. Experimental results across multiple MEP systems demonstrate that the proposed approach achieves high segmentation accuracy and computational efficiency, without relying on large-scale annotated instance training data.