基于计算机视觉的钢桥螺栓松动检测方法

Computer Vision-based Detection Method for Steel Bridge Bolt-looseness

为提高螺栓松动检测的智能化水平,提出一种基于计算机视觉的钢桥螺栓松动检测方法。首先基于深度学习理论建立关键点检测模型,对采集的螺栓图像进行标注并建立数据集;然后分别训练目标检测模型YoloV5和关键点检测模型,并利用训练后的模型自上而下检测螺栓关键点,根据关键点确定螺栓中心点位置,以中心点的相对位置求解透视变换矩阵,利用透视变换矩阵对关键点进行重投影;最后根据关键点的位置变化检测螺栓是否发生松动。结果表明:训练后的YoloV5模型和关键点检测模型可准确检测出螺栓的关键点;关键点的检测精度受图像采集条件影响且对角度更为敏感;利用所有中心点拟合透视变换矩阵的最小二乘解可提高图像几何矫正的精度;不同图像采集环境下,松动螺栓的检测误差在0%~9.6%之间,误检率为2.7%,表明本方法的检测精度和稳定性均较高,具有较好的实用价值和广阔的工程应用前景。

In order to improve the intelligence of bolt-looseness detection,a computer vision-based detection method was proposed for steel bridge bolt-looseness.Firstly,bolt keypoint detection model was established based on deep learning theory to annotate the collected bolt images and to build datasets.Then the object detection model YoloV5 and the keypoint detection model were trained separately to detect the bolt keypoints from top to bottom using the trained models.The location of bolt center points was determined according to the keypoints,and the perspective transformation matrix was solved according to the relative position of the center points,which was then used to reproject the keypoints.Finally,bolt-looseness was detected according to the position changes of keypoints.The results show that the trained YoloV5 model and keypoint detection model can accurately detect the keypoints of the bolts.The detection accuracy of the keypoints is affected by the image acquisition conditions and is more sensitive to angles.Fitting the least-squares solution of the perspective transformation matrix using all center points can improve the accuracy of image geometry correction.The detection error of bolt-looseness under different image acquisition conditions ranges from 0%to 9.6%,with a false detection rate of 2.7%,indicating that the proposed method,with high accuracy and stability,has great practical value and broad engineering application prospects.