摘要
近年来,基于深度学习的缺陷销钉检测技术在电力杆塔巡检中得到广泛应用。但是该技术只能在图像空间对缺陷销钉进行定位,无法准确给出缺陷销钉在实际电力杆塔中对应销钉的唯一编号。针对该问题,该文提出一种融合深度学习与三维重投影的缺陷销钉唯一性识别方法。首先通过构建巡检电力杆塔系统的三维模型,进行电力杆塔系统中部件模型的编号获取,以及部件模型三维空间包围盒顶点和中心点坐标的计算;其次,利用YOLOv5模型在图像空间对销钉及上下文相关目标进行检测;再次,通过该文提出的融合结构约束的相机位姿估计算法对航拍图像的相机位姿进行估计;最后,将三维电力杆塔系统中的销钉包围盒按相机位姿进行重投影与销钉检测框进行匹配度计算,在图像空间中输出带有具体编号的销钉检测框。在YOLOv5模型输出的销钉检测框基础上进行销钉唯一性识别,利用仿真数据进行实验得到的唯一性识别正确率为100%,利用实拍数据进行实验得到的唯一性识别正确率为93.3%,验证了该文提出的销钉唯一性识别方法的有效性。通过对缺陷销钉进行唯一性识别,可以减少人工对故障图像的二次审核工作量,并准确地统计缺陷销钉的数量及具体位置信息,为后续故障维修及故障相关性分析等精细化管理提供支持。
With the development of artificial intelligence technology,defect pin detection technology based on deep learning has been widely applied in the field of power tower inspection.However,this technology can only locate the defect pin in the image space,but cannot provide a unique number for the defect pin in the actual power tower,and cannot automatically determine the actual position of the defect pin.Moreover,for multi-angle images of the same part,the defect pin detection algorithm using depth learning may repeatedly label the same defect pin,which may cause repeated statistics of defect pins.To address this issue,this paper proposes a pin uniqueness identification method that integrates deep learning and 3D reprojection,which is essentially an extension of the 2D object detection task.Firstly,a three-dimensional model of the inspection power tower system is constructed to obtain vertices and center points of the components 3D space bounding box,and unique number of the components in the power tower system.Then,the YOLOv5 model is used to detect pins and contextual sensitive targets in the image space.The contextual targets include a group of large-scale hardware,a mesoscale specific hardware,and a small-scale bolt head.Filter detection results that are not within the inspection area by using the hardware target group bounding box.Then use the camera pose estimation algorithm that integrates structural constraints proposed in this article to estimate the camera pose of the image.Use the center points of the pins and bolt heads 2D bounding boxes as 2D feature points,and the center point of the pins and bolt heads 3D space bounding boxes as 3D feature point to form point-to-point constraints.Use P3P algorithm for camera pose estimation.For the unknown 2D-3D correspondence,four 2D feature points are selected by using the specific hardware bounding box,and the matching solution space of 2D-3D feature points is constructed by enumeration method.Calculate the camera pose for each solution separately,and use the deviat
作者
阎光伟
马颐琳
焦润海
何慧
Yan Guangwei;Ma Yilin;Jiao Runhai;He Hui(School of Control and Computer Engineering North China Electric Power University,Beijing 100096 China)
出处
《电工技术学报》
EI
CSCD
北大核心
2024年第17期5450-5460,共11页
Transactions of China Electrotechnical Society
基金
国家自然科学基金资助项目(62073133)。
关键词
电力巡检
销钉检测
相机位姿估计
唯一性识别
Power inspection
pin detection
camera pose estimation
uniqueness identification
