Deep learning-based 3D digital damage map of vertical-type tunnels using unmanned fusion data scanning

Abstract

Expert-dependent visual inspection of vertical-type tunnels is often not only dangerous for workers, but also unreliable of the inspection results. To address the technical issues, a deep learning-based 3D digital damage mapping technique using an unmanned fusion data-scanning system is described in this article. The proposed system acquires raw digital images and point cloud data by moving vertically within the tunnel. Subsequently, optimal images are automatically selected and used for deep learning-based damage evaluation and fixed camera pose-based 3D digital modeling. Finally, a 3D digital damage map is generated by mapping the damage features onto the 3D digital map based on the predetermined camera pose. The proposed technique was experimentally validated in two different in-situ testbeds in South Korea, revealing that averaged precision and recall values are 90.89% and 98.18%, respectively. It is expected that the proposed technique can improve in-situ workability as well as data reliability.