Unsupervised seepage segmentation pipeline based on point cloud projection with large vision model

Abstract

Targeting to sustain the operational safety of tunnels, this paper presents a projection-based approach utilizing large vision model (LVM) inference for efficient and effective multi-class segmentation, including the identification of seepage, in 3D point clouds of tunnels. The proposed approach employs an unsupervised strategy based on point projection and label correction to process input point clouds and enhance semantic inference, particularly in distinguishing between seepage and other segment classes. Furthermore, a large vision model is adopted to improve the method with practical application. To assess the method’s effectiveness, real tunnel point cloud data from a cross-river tunnel section in China is utilized. The results demonstrate: (1) The proposed framework excels in unsupervised seepage segmentation, achieving F1 and Accuracy scores of 0.769 and 0.930, respectively; (2) Competitive unsupervised detection results are achieved across multiple classes, with segmentation F1 scores of 0.743, 0.841, 0.983, 0.970, 0.836, and 0.933 for cable, segment, pipe, powertrack, support, and track classes, respectively; (3) The segmentation model outperforms other state-of-the-art unsupervised point cloud segmentation algorithms and demonstrates competitive performance compared to supervised methods. Overall, this developed method holds promise as the foundation for an automated decision-making system, facilitating tunnel surveys and improving segmentation effectiveness.