From pixel to infrastructure: Photogrammetry-based tunnel crack digitalization and documentation method using deep learning

Crack detection and documentation play a vital role in the asset management of large-scale tunnel complexes. This study proposes a computer vision-based tunnel crack data management method enabling 3D visualization, quantification, and documentation into structured data. The method reconstructs sparse point clouds with Structure from Motion (SfM) and cleans the irrelevant tunnel facilities with a two-stage filtering method. The denoised 3D point clouds are then fitted with customised meshes and textured into 3D reconstruction models. The flat scaled orthomosaic is generated by the cylindrical unrolling. Deep learning methods are employed for pixel-level crack detection in this high-resolution image for the extraction of crack location and quantification. Applied to four tunnel sections of CERN, the European Organization for Nuclear Research, the method presents the spatial crack distributions and quantifies crack dimensions. In addition, the original unstructured tunnel image data of 1024 MB / meter is converted to structured tabular data of 0.3 MB / meter. The quantification result provides a crack statistical analysis tool, revealing that the crack length meets the log-normal distribution indicating the inherent fractural characteristic of tunnel linings.