Noncontact Structural Vibration Displacement Measurement Using Deep Learning–Based Optical Flow Estimation

In contrast with traditional procedures, computer vision–based approaches to the noncontact measurement of structural vibration displacement offer the benefits of convenience, efficiency, and dense sensing. These approaches provide comprehensive noncontact detection by using the natural characteristics of the structure. Although current research offers various approaches verified by indoor experiments or small-scale engineering testing, there remains a deficiency in considerations for practical application. Performance under varying vibration amplitudes has not been quantified, and discrepancies among various approaches lack standardized evaluation. The capacity to resist bad weather and dramatic illumination change has not been adequately investigated, constraining practical application. This paper investigates a visual sensing approach for noncontact structural vibration displacement measurement based on deep optical flow inference. Indoor experiments with different vibration amplitudes were conducted and comparisons made simultaneously with three other typical deep learning–based approaches to demonstrate that the approach has the widest identification range and highest accuracy. Based on a four-floor frame model experiment data, the approach is illustrated to offer the highest resistance to bad weather and dramatic illumination change by manually adding environmental disturbances and then be further verified via environmental disturbances experiments. The approach was ultimately utilized for the frequency identification of hangers on a kilometer-level suspension bridge, yielding results that accurately identified the first several orders of frequency for three different lengths, with relative errors not exceeding 0.4% in comparison with accelerometers. This approach shows superior identification accuracy and resistance to environmental disturbances, which can provide technical support for practical engineering.