Tunnel crack assessment using simultaneous localization and mapping (SLAM) and deep learning segmentation

Abstract

Artificial intelligence algorithms and multi-sensor technologies are advancing tunnel crack detection. However, image-based detection methods fail to account for tunnel section curvature, limiting their ability to represent the spatial geometry of cracks. To address these problems, this paper presents a tunnel crack assessment method combining simultaneous localization and mapping (SLAM) with deep learning-based segmentation. The SLAM algorithm reconstructs the tunnel point cloud map, and a two-dimensional (2D) convex hull point cloud unfolding with a cloth simulation filter (CSF) algorithm is applied for denoising. A deep learning segmentation model is used to segment the tunnel cracks. The cracks are projected into a three-dimensional (3D) point cloud map, and the crack length and spatial location are calculated. Field tests demonstrate that the method reduces tunnel reconstruction time to 27 s (a 99 % time saving), with a maximum radius error of 0.03 m and accurate 3D crack projections.