基于方向自适应检测器的输电线路设备检测方法  

Oriented Object Detector for Transmission Line Equipment Detection

作　　者：刘海莹 莫文昊 谈元鹏 刘佳鑫 李勇 LIU Haiying;MO Wenhao;TAN Yuanpeng;LIU Jiaxin;LI Yong(China Electric Power Research Institute,Haidian District,Beijing 100192,China;State Grid Liaoning Electric Power Research Institutet,Shenyang 110000,Liaoning Province,China;State Grid Shandong Electric Power Company,Jinan 250001,Shandong Province,China)

机构地区：[1]中国电力科学研究院有限公司,北京市海淀区100192 [2]国网辽宁省电力有限公司电力科学研究院,辽宁省沈阳市110000 [3]国网山东省电力公司,山东省济南市250001

出　　处：《电网技术》2021年第12期4888-4895,共8页Power System Technology

基　　金：国家重点研发计划项目(2018YFB1307400);国家电网有限公司科技项目(SGSDDK00KJJS2000090)。

摘　　要：如何利用现有可见光影像数据与设备,实现高效巡检,辅助一线作业人员开展工作,是目前电力自动化巡检研究中亟待解决的难题。基于可见光图像的电力巡检存在图像畸变、待检测物体和摄像机角度不同导致目标特征丢失等问题,常见的目标检测算法往往效果较差,无法满足电力巡检要求。针对上述问题,提出一种基于CenterNet的有向检测器Rot-CenterNet。具体方案:首先,为了检测有向目标框,加入用于回归角度的检测头,并引入IoU-L1计算目标检测头的损失函数。其次,Rot-CenterNet提出3个骨干网络以适应于不同算力的电力业务场景部署,分别为保持高分辨率表征的HRNet、参数量少且实现精度与速度极致性价比的EfficientNet和大多数边缘芯片均支持的经典算子ResNet。同时,该文设计了DCN-ASPP和D-SKN模块,实现感受野随目标设备的形状和角度方向自动调整。最后,针对现有输电线路可见光数据集较少且不规范的问题,以项目为依托,整理了一批包括架空输电和电缆隧道场景在内的有向设备数据集并命名为TransLine-2020。在测试集上,经过检测器和骨干网络的改进,所提出的模型在检测设备元件上,相比CenterNet模型平均精度值(average percision,AP)提高了5.95。为了进一步证明检测器具备多场景应用能力,Rot-CenterNet在公开DOTA数据集中也进行了实验,取得了同样不错的效果。How to better employ the present visible-light based inspection equipment to assist the front-line workers to carry out more efficient transmission line inspection routine is the primary need to be satisfied in practice.The captured visible-light inspection images are often accompanied with image distortion as well as large inclination angles between the objects and the cameras,which pose challenges for most present application level object detection algorithms.To solve the above problems,based on the improved CenterNet,this paper presents an oriented object detection network-the Rot-CenterNet.First,to detect the oriented object bounding box,we append extract detection head for regressing the angle value and introduce the improved IoU-L1 to calculate its loss.Second,Rot-CenterNet provided three backbone networks to fit different computing power application scenarios,namely HRNet with high resolution representation throughout the network,EfficientNet which achieved optimal balance between speed and accuracy,and classic operator ResNet supported by most edge devices.At the same time,this paper designs DCN-ASPP and D-SKN modules to implement the adaptive adjustment of the receptive field according to shape and angle direction of the target device ground truth.Last,to address the limited availability of the related benchmarks on the transmission line inspection,based on the project,we collect an extensive and fully annotated oriented dataset including overhead transmission line and cable tunnel scenarios,namely,TransLine-2020.On the test set,with improvements in the detector and backbone network structure,the proposed model improved 5.95 compared to the CenterNet model Average Percision value(AP).In addition,the Rot-CenterNet has also performed well on the remote sensing dataset DOTA,further proving the generalization ability of the detector.

关 键 词：输电线路 电缆隧道 架空输电 深度学习 有向目标检测器 

分 类 号：TM721[电气工程—电力系统及自动化]