A hybrid approach for vision-based structural displacement measurement using transforming model prediction and KLT

Abstract

As civil infrastructures age, monitoring their health conditions has become increasingly critical. Dynamic displacement measurement is a prevalent method for assessing structural health. Traditional techniques, which often involve installing instruments and scaffolding, can interfere with the response of the structures. To address the challenge, non-contact measurement methods have been developed; however, these are typically costly and require expert operation. Advances in high-speed industrial cameras and image processing technology now enable vision-based displacement measurement. Despite their effectiveness, existing vision-based methods face significant limitations, including their inability to maintain tracking when line-of-sight is obstructed, sensitivity to lighting variations, and the need for manual intervention when feature points are lost. This study introduces a novel hybrid approach, termed ToMP-KLT, which combines the KLT tracker with a deep learning-based model. This method harnesses the precision of the KLT tracker under favorable conditions and the robustness of the deep learning-based tracker under adverse conditions. Its effectiveness is validated through simulation-based tests, lab-scale experiments, and field testing on Cheonsa Bridge, demonstrating substantial improvements in tracking robustness against occlusions and varying light conditions.