SAM4Tun: No-training model for tunnel lining point cloud component segmentation

Abstract

Asset management ensures the safety and longevity of structures through regular maintenance. Reality capture technologies are increasingly being used for asset inspections to obtain information by generating point cloud data, which is becoming more prevalent in tunnel asset management for precise documentation of tunnel geometry and condition. Integrating semantic information from point clouds is crucial for creating accurate as-built Building Information Models (BIM), essential for project delivery, maintenance, and operations. In this paper, we propose SAM4Tun, a zero-shot automated instance segmentation method for tunnel lining segments. It is based on a Large Vision Model (LVM), prompt-based Segment Anything Model (SAM), and various point cloud and image processing techniques, enabling accurate instance segmentation without requiring any training. The process starts by unfolding tunnel point clouds to generate 2D panoramic images, enabling SAM to be extend its capabilities to point cloud segmentation. To enhance performance, we propose: (i) a local point cloud density-variation method to filter out non-segment parts, and (ii) a geometry feature-guided multi-step point cloud up-sampling method to address uneven point cloud density during projection. Then, we focus on prompt engineering, using traditional image processing techniques to automatically generate template prompt, enabling SAM’s zero-shot ability to achieve precise instance-level segmentation of tunnel linings. The results demonstrate that our no-training model achieved highly accurate instance segmentation, even surpassing supervised learning algorithms. The proposed method addresses the issue of data dependency and serves as the foundation for component-level damage localization and displacement monitoring in tunnel. Our code is available at https://github.com/zxy239/SAM4Tun.