An anti-occlusion vision-based method for structural motion estimation

Structural displacement is an important metric in structural health monitoring (SHM). Computer vision-based methods have been widely used for structural displacement recognition in laboratory settings. However, the movement of natural objects such as pedestrians, vehicles, and other unrelated objects, may obstruct the selected structural targets for tracking, therefore reducing the accuracy of estimated displacements. To address this challenge, this paper proposes an anti- occlusion computer vision-based method to estimate structural displacements with subpixellevel accuracy. The proposed method can be divided into three steps. First, the correlation filter is used to continuously track the selected target despite occlusion. Next, the Gaussian mixture model (GMM)-based target modeling method is proposed to identify the occluded segments in each frame. Finally, the selected target is divided into multiple patches, and the subpixel-level displacements of unobstructed patches are estimated by the subpixel patch matching algorithm. The advantages of the developed method over traditional approaches are demonstrated in simulated cases and a shaking table test of a five-story stone curtain wall structure. Additionally, the developed method is applied in a bridge construction practice.