Deep learning-based pipe segmentation and geometric reconstruction from poorly scanned point clouds using BIM-driven data alignment

Abstract

Pipe reconstruction is an important prerequisite for pipe maintenance. However, scanned point clouds often contain defects, presenting a significant challenge for automated segmentation and geometric reconstruction. To address this challenge, this paper proposes a learning-based segmentation method, PipeSegNet, along with a geometric reconstruction process. In the segmentation stage, a method is developed to generate datasets with controlled density from BIM. Meanwhile, alignment strategies are introduced to address feature and label inconsistencies between BIM-generated and real datasets. PipeSegNet enhances global and local perceptual capability, achieving pipe segmentation accuracy of 96.37 % and IoU of 91.45 %, ensuring high-quality reconstruction. Comparative and module evaluation experiments demonstrate the effectiveness of PipeSegNet combined with the alignment strategies. The total average relative error of the reconstructed pipes is 2.73 %. This paper provides valuable insights into the pipe segmentation and reconstruction from point clouds, particularly in scenes with poor scanning quality, contributing to efficient infrastructure maintenance.