Structural displacement is an important quantity to assess the health of civil infrastructure. Vision‐based approaches using unmanned aerial vehicles (UAV) mounted with high‐resolution cameras have been proposed for this purpose. However, because the camera itself is moving with the UAV, any video obtained will contain both the motion of the structure and the motion of the camera. Planar homography can be used to eliminate the errors induced by the camera movement without the need for camera parameters. However, its direct application to large structures still has limitations, because capturing the undeformed regions, along with the measurement points on the structure, within a single image with sufficient resolution is seldom feasible. In this study, a new framework is presented to address these issues and facilitate the extraction of the structural displacement from videos taken by a UAV‐mounted camera. First, a two‐layer feedforward neural network (FNN) is adopted to obtain the image coordinates of the selected features of the structure on its stationary position, which are further used as homography features. Next, the structural displacement is estimated with the homography transformation matrix determined from the obtained homography features. Finally, the proposed approach is validated on both a six‐story shear‐building model in the laboratory and an elevator tower located in Zhongshan City, China. These results demonstrate the efficacy of the proposed approach.

Structural vibration measurement is a crucial and necessary step for structural health monitoring. Recently, computer vision-based techniques have been proposed by researchers to measure structural motion remotely. However, the direct application of vision-based measurement to practical applications still faces some challenges, mainly because intrinsic camera vibration can introduce significant errors to the measurement results. In this study, a three-stage approach using an embedded inertial measurement unit is proposed to compensate for the camera movement. First, camera rotations are estimated by employing a complementary filter with an adaptive gain to fuse gyroscope measurement and accelerometer data. Next, binary robust invariant scalable key-point features are detected from the region of interest and tracked between video frames using a Kanade-Lucas-Tomasi tracker. Finally, structural acceleration is obtained by combining the information for the obtained structural features and the estimated nonstationary camera motion. The performance of the proposed approach is investigated using both a moving handheld camera and a camera mounted on the unmanned aerial vehicle in the laboratory. These results demonstrate that the proposed method can be effectively applied to measure structural vibration, without requiring stationary background features to be available in the video.

The dynamic measurement and identification of structural deformation are essential for structural health monitoring. Traditional contact-type displacement monitoring inevitably requires the arrangement of measurement points on physical structures and the setting of stable reference systems, which limits the application of dynamic displacement measurement of structures in practice. Computer vision-based structural displacement monitoring has the characteristics of non-contact measurement, simple installation, and relatively low cost. However, the existing displacement identification methods are still influenced by lighting conditions, image resolution, and shooting-rate, which limits engineering applications. This paper presents a data fusion method for contact acceleration monitoring and non-contact displacement recognition, utilizing the high dynamic sampling rate of traditional contact acceleration sensors. It establishes and validates an accurate estimation method for dynamic deformation states. The structural displacement is obtained by combining an improved KLT algorithm and asynchronous multi-rate Kalman filtering. The results show that the presented method can help improve the displacement sampling rate and collect high-frequency vibration information compared with only the vision measurement technique. The normalized root mean square error is less than 2% for the proposed method.