Tracking vibration displacement across multiple points on large-scale bridges in the field poses significant challenges due to perspective distortion by camera tilting, interference from adverse environmental factors and low-resolution images by long distance measurement. To address these issues, this study proposed a novel vision-based displacement measurement framework that combined the Fast Spatio-Temporal Context Learning algorithm (FSTC) with tilt shift cameras to enhance visual tracking accuracy. A procedure was introduced to select suitable tilt shift cameras to eliminate the adverse effects of potential lens distortion. The FSTC algorithm, which integrated the Spatio-Temporal Context Learning algorithm (STC) and FastFlownet, was employed to improve the displacement tracking accuracy of the target object under fog occlusion and during long-distance measurements. Experimental results from slider and shaking table tests demonstrated that the FSTC algorithm outperformed the existing vision-based methods such as STC, Digital Image Correlation (DIC), and Kanade-Lucas-Tomasi (KLT) algorithms in tracking the displacement of the target object, effectively controlling the displacement error (e.g., RMSE and peak error) to below 3 mm. The FSTC algorithm was less sensitive to fog occlusion and exhibited enhanced efficiency in object tracking. Furthermore, the FSTC algorithm, in combination with image resolution reconstruction, enabled subpixel accuracy in tracking the displacement of the target object. A preliminary field application was conducted to track the vibration displacement of the Sanchaji Bridge under wind load, validating the feasibility and applicability of the proposed framework.

中文翻译：

---

基于视觉的基于倾斜移位相机和快速时空上下文学习的大型桥梁位移测量方法


由于摄像机倾斜造成的视角失真、不利环境因素的干扰以及远距离测量的低分辨率图像，在现场跟踪大型桥梁上多个点的振动位移带来了重大挑战。为了解决这些问题，本研究提出了一种新的基于视觉的位移测量框架，该框架将快速时空上下文学习算法 （FSTC） 与移轴相机相结合，以提高视觉跟踪精度。引入了一种程序来选择合适的移轴相机，以消除潜在镜头失真的不利影响。采用融合了时空上下文学习算法 （STC） 和 FastFlownet 的 FSTC 算法，提高了雾遮挡和远距离测量时目标物体的位移跟踪精度。滑块和振动台测试的实验结果表明，FSTC 算法在跟踪目标物体的位移方面优于现有的基于视觉的方法，如 STC、数字图像相关 （DIC） 和 Kanade-Lucas-Tomasi （KLT） 算法，有效地将位移误差 （例如 RMSE 和峰值误差） 控制在 3 mm 以下。FSTC 算法对雾遮挡不太敏感，并且在目标跟踪方面表现出更高的效率。此外，FSTC 算法与图像分辨率重建相结合，在跟踪目标物体的位移时实现了亚像素精度。进行了初步的现场应用，以跟踪三姑寺大桥在风荷载下的振动位移，验证了所提出的框架的可行性和适用性。