Monocular vision-based structural out-of-plane motion estimation using a deep learning method

Abstract

Computer vision-based displacement estimation methods can estimate structural motion from videos, but face challenges in estimating out-of-plane motion. Conventional vision-based methods employ either multi-view camera systems or structural model-aided monocular vision methods to estimate structural out-of-plane motion. Nonetheless, the utilization of multi-view camera systems leads to additional costs and stringent requirements for overlapping views, while structural model-aided methods require comprehensive structural information. To achieve convenient and accurate out-of-plane structural displacement measurement, this study presented a monocular vision-based measurement method using a deep learning technique. Firstly, a deep learning model was employed to estimate the global depth matrix of images. The relative depth of the region of interest (ROI) was then estimated through pyramid-accelerated tracking and relative depth matrix weighting. To convert the relative depth into actual out-of-plane motion, the scale factor was calculated through least squares fit. The performance of the proposed method was validated through static and dynamic tests. The results indicated that the proposed approach yielded reasonable out-of-plane motion estimations, with correlation coefficients with displacement sensor measurements over 0.93. The proposed method therefore offered a cost-effective and convenient alternative for structural out-of-plane motion estimation.