Enhanced vision-based structural displacement monitoring through deep learning approaches

Vision-based displacement monitoring for large-scale civil infrastructures remains challenged by limited imaging resolution and uncontrolled camera-motion. This study presents a hybrid deep learning (DL) framework addressing these dual challenges through two technical innovations. Firstly, we develop an enhanced video super-resolution (VSR) architecture based on BasicVSR++, incorporating a novel multi-scale feature extraction module with pre-alignment mechanism which uses a multi-stage bidirectional propagation strategy to optimize temporal feature fusion. Secondly, we devise a dual-stage convolutional neural networks (CNN) architecture for unsupervised homography (H) estimation, enabling coarse-to-fine camera motion compensation through parametric transformation. The integrated displacement measurement method combines super-resolved imagery with KAZE-DIC algorithm for sub-pixel target tracking under challenging conditions including low illumination, texture-deficient backgrounds, and camera-motion. Field validation on an 888-meter suspension bridge demonstrates the framework's potential for structural health monitoring applications. The proposed methodology advances vision-based metrology by simultaneously resolving resolution constraints and motion artifacts through synergistic DL strategies.