DISTRESS DETECTION BASED ON VISION TRANSFORMER USING EGOCENTRIC VIDEOS WHILE INSPECTING IN SUBWEY TUNNELS

Abstract

In this paper, we propose a distress detection method based on Vision Transformer using egocentric videos of engineers inspecting in subway tunnels. The proposed method detects distresses in subway tunnels by fine-tuned Vision Transformer trained by a large-scale image dataset. Furthermore, we train Vision Transformer by DINO, a self-supervised learning method, to generate an attention map that can be used as a reason for the distress detection result. As a result, the proposed method achieves highly accurate detection of distresses in subway tunnels and can provide a reason for the distress detection results. In the last part of this paper, we confirm the effectiveness of the proposed method by the experiment using actual egocentric videos of engineers inspecting in subway tunnels.