This study proposesStructure-PoseNet to recognize three-dimensional (3D) poses and dense dynamic displacement of a structure with known geometry using a monocular camera. The 3D poses are employed to calculate the structural displacement of both visible and invisible elements in videos. The proposed framework consists of two consecutive deep learning modules, CompNet and ParaNet, to provide semantic image segmentation and extract pose parameters, respectively. CompNet converts natural video images into semantic classification masks. ParaNet uses the recognized masks from CompNet as input and outputs the pose parameters. The dense dynamic displacement at measurement points with an intervaldistance of 2% of the size of the structure (along the direction with the maximum size) is then obtained using the coordinates of each 3D model vertex. The identification accuracy is discussed for different cases. Furthermore, the performance of the proposed method was validated through the shaking table test results of the two scale models. Finally, the proposed method was applied to a full-scale suspension bridge and the Tacoma Narrows Bridge (1940) using Internet videos.

中文翻译：

---

Structure-PoseNet，用于使用单目相机识别结构的密集动态位移和三维姿态

这项研究提出了Structure-PoseNet，以使用单目相机识别具有已知几何形状的结构的三维（3D）姿势和密集动态位移。3D 姿势用于计算视频中可见和不可见元素的结构位移。所提出的框架由两个连续的深度学习模块 CompNet 和 ParaNet 组成，分别提供语义图像分割和提取姿势参数。CompNet 将自然视频图像转换为语义分类掩码。ParaNet 使用来自 CompNet 的识别掩码作为输入并输出位姿参数。然后使用每个 3D 模型顶点的坐标获得在间隔距离为结构尺寸 2% 的测量点处（沿最大尺寸方向）的密集动态位移。针对不同情况讨论识别精度。此外，通过两个比例模型的振动台测试结果验证了所提方法的性能。最后，使用互联网视频将所提出的方法应用于全尺寸悬索桥和塔科马海峡大桥（1940）。