Deep learning-based motion magnification and frames matching for structural displacement measurement using computer vision

Compared with traditional contact and non-contact sensors used for acquiring vibration displacement responses, fixed cameras with the assistance of computer vision (CV) technology, offer lower costs and a greater flexibility in structural health monitoring (SHM). However, capturing subtle structural vibrations remains challenging due to low motion visibility and visual degradation caused by low illumination, weak textures, and image noise. To overcome this challenge, this paper proposes a vision-based structural displacement monitoring approach comprising three main steps, namely, camera calibration, subtle motion magnification, and frames matching for estimating displacement. The state-of-the-art deep learning-based motion magnification method (EulerMormer) is employed to improve the visibility of these subtle vibrations and exhibit fewer artifacts and distortions without prior knowledge of the target structure. A hybrid frames matching method termed Phase-ECC is designed, integrating phase correlation and enhanced correlation coefficient (ECC) to robustly estimate sub-pixel inter-frame translations across a wide range of vibration magnitudes. The proposed approach is validated through a rigid beam under controlled excitations, a cantilever beam excited by a shaker, and a vehicle-bridge interaction scenario under both normal and low-light conditions. The results demonstrate that the applied EulerMormer offers a significant advancement in minimizing distortion and artifacts, especially under large amplification factors or large-magnitude vibrations compared to other existing methods. The proposed approach accurately measures micro structural displacements, outperforming traditional vision-based techniques in terms of tracking accuracy, noise robustness, and low-light performance. • This paper proposes a novel structural displacement measurement approach for SHM. • The proposed approach is built on deep learning-based motion magnification. • Motion magnification and frames matching are used to process vision videos. • The effectiveness and accuracy of the proposed approach are demonstrated. • The proposed approach is verified with experimental studies and modal identification.