Deep learning-based tunnel surface defects 3D reconstruction and spatial location mapping method

Abstract

Spatial location information of tunnel surface defects is crucial for tunnel safety assessment. Existing methods mainly focus on 2D defect detection and lack efficient 3D spatial localization capability. To address the issue, a deep learning-based tunnel surface defect 3D reconstruction and spatial mapping method was proposed in this paper. The method first selects valid images through a keyframe extraction strategy, then utilizes ResNext-SKnet to extract multi-scale features. It combines voxel conversion with GRU recurrent fusion to achieve incremental 3D modeling, and finally maps defect information onto the 3D model via MVS-Texturing. Indoor tunnel model comparison tests demonstrate that this method achieves 494,000 reconstructed points, 708,982 triangular surfaces, and a reconstruction frame rate of 29 fps. Field tests reveal an average damage size error of only 1.44%, meeting practical engineering requirements.