基于计算机视觉和深度学习的隔震支座动态位移测量方法  

Dynamic displacement measurement method for an isolation bearing based on computer vision and deep learning

作　　者：党育[1] 贺一哲 DANG Yu;HE Yizhe(School of Civil Engineering,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学土木学院,兰州730050

出　　处：《振动与冲击》2023年第6期90-97,165,共9页Journal of Vibration and Shock

基　　金：国家自然科学基金(62166025,51668043);甘肃省重点研发计划(21YF5GA073)。

摘　　要：针对地震时隔震支座的视觉测量特点,即地震时相机与隔震支座均会发生运动,且不便追踪相机位姿,提出一种无需额外跟踪相机运动的隔震支座动态位移测量方法。首先,沿结构水平两个正交方向,分别设置相机和标靶,且相机光轴与标靶平面相互垂直,用比例因子法对各方向各时刻的相机进行标定;其次,采用深度学习方法,训练一个YOLOv5模型以自动识别和定位标靶,再对YOLOv5的目标框放大并对目标框内的图像进行特征识别和提取,二次精确定位标靶;最后,使用支座上、下连接板处标靶的相对位移,确定隔震支座的动态位移。通过拍摄一个LRB500隔震支座力学性能试验过程,对提出的方法进行验证。结果表明:该方法得到的支座水平位移时程曲线与位移计结果在各时刻的误差均小于1.0%,水平向位移峰值的绝对误差最大为1.042 mm,竖向位移峰值的绝对误差最大为0.219 mm,说明提出的方法具有较高的检测精度,可以在无需追踪相机位姿的情况下,完成隔震支座在地震作用下的变形检测。The visual measurement characteristics of an isolation bearing during earthquake,was aimed at,because of considering the fact that both the camera and the isolation bearing will move during earthquake,and it is inconvenient to track the camera pose.A dynamic displacement measurement method for the isolation bearing without the netessity of additional tracking of the camera movement was proposed.Firstly,the camera and the target were set in orthogonal horizontal directions with the optical axis of the camera and the target plane being perpendicular to each other,and the scale factor method was used to calibrate the camera in each direction at each time.Then,using the deep learning method,a YOLOv5 model was trained to automatically identify and locate the target,moreover,the prediction box of YOLOv5 was enlarged and the target was repositioned by performing feature recognition and extraction on the enlarged image.Finally,the relative displacement of the target at the upper and lower connecting plates of the bearing was used to determine the dynamic displacement of the isolation bearing.The method was verified by photographing the mechanical performance test process of a LRB500 isolation bearing.The results show that the error between the horizontal displacement time-history curve of the bearing obtained by this method and the displacement meter results at each moment is less than 1.0%,the absolute error of the horizontal displacement peak value is 1.042 mm,and the absolute error of the vertical displacement peak value is 0.219 mm.It is shown that the method proposed has high detection accuracy,and can complete the deformation detection of the isolation bearing under the action of earthquake without tracking the camera pose.

关 键 词：深度学习 计算机视觉 隔震支座 动态位移 YOLOv5 

分 类 号：TU352.1[建筑科学—结构工程]