Deep Learning-Based Rapid 3D Reconstruction and Spatial Localization of Leakage in Shield Tunnels

To address the inefficiencies and limitations of traditional three-dimensional (3D) reconstruction methods for shield tunnels, particularly in visualization and leakage localization, this study proposes a deep learning–based framework that combines monocular depth estimation networks with truncated signed distance function (TSDF) technology for rapid 3D point cloud reconstruction. A joint supervised training strategy is employed to overcome the limitations of self-supervised methods that rely on reprojection error, particularly in structurally similar shield tunnel environments. This strategy significantly enhances depth estimation performance. Additionally, a pose optimization method based on feature point spatial density is introduced to mitigate geometric deviations caused by pose errors. Experimental results demonstrate that the proposed framework achieves a geometric accuracy error of only 1.5%, while requiring just 2.3% of the modeling time compared to traditional SfM-based methods. By integrating with image segmentation networks, the framework allows for precise spatial localization of leakage defects. The rapid 3D reconstruction of shield tunnels equips monitoring personnel with timely spatial information on leakage defects, facilitating effective and prompt interventions.