Vision‐based displacement measurement enhanced by super‐resolution using generative adversarial networks

Abstract

Monitoring the deformation or displacement response of buildings is critical for structural safety. Recently, the development of computer vision has led to extensive research on the application of vision‐based measurements in the structural monitoring. This enables the use of urban surveillance video cameras, which are widely installed and can produce numerous images and videos of urban scenes to measure the structural displacement. However, the structural displacement measurement may be inaccurate owing to the limited hardware resolution of the surveillance video cameras or the long distance from the cameras to the monitored targets. To this end, this study proposes a method to improve the displacement measurement accuracy using a deep learning super‐resolution model based on generative adversarial networks. The proposed method achieves texture detail enhancement of low‐resolution images or videos by supplementing high‐resolution photographs of the target, thus improving the accuracy of the vision‐based displacement measurement. The proposed method shows good accuracy and stability in both the static and dynamic experimental validations compared with the original low‐resolution images/video and interpolation‐based super‐resolution images/video. In conclusion, the proposed method can support the displacement measurement of buildings and infrastructures based on urban surveillance video cameras.