输电线路图像上防震锤检测算法研究

发布时间：2018-03-19 05:40

本文选题：输电线提取　切入点：防震锤检测　出处：《北京交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：在智能电网系统中,输电线路的智能巡检是极其重要的组成部分。输电线路上的防震锤是一种为了减少输电线因风力扯起振动而安装的设备,它是保障电力正常运输的重要一环。因此,对输电线路图像上防震锤检测的研究就彰显出重要的价值。传统的输电线路的巡检工作基本上采用人工巡检的方式,这种作业方式劳动强度大、危险系数高并且效率低。近年来,无人机航拍的巡检方式被部分电力公司使用。然而,无人机采集到的图像视野辽阔、背景复杂、视角、目标分辨率也不尽相同,这些都给防震锤检测带来了极大的挑战。本论文正是基于这一背景下,调研国内外相关工作,对输电线路上防震锤的检测定位算法展开研究,以解决复杂背景中鲁棒的检测防震锤的问题,为防震锤的缺陷诊断做准备。本文的主要工作有:(1)基于AdaBoost算法的防震锤检测。针对现有的基于随机霍夫变换、基于模板匹配的方法检测,只适用于特定场景、特定角度的防震锤检测,并不适用于本数据集的问题。本文提出了一种基于AdaBoost算法的防震锤检测方法。在训练阶段,我们使用了针对防震锤的Haar特征然后结合AdaBoost分类器进行训练。在检测阶段,由于防震锤与输电线的依附关系,我们首先对输电线进行提取,候选出带有防震锤的感兴趣区域,然后通过滑动窗口结合分类器在此区域进行防震锤的检测。最后我们比较分析了不同的特征选取、不同的参数设置对实验结果的影响,并从准确率和计算效率两方面对实验结果进行评价与分析。在无人机采集的野外的输电线路图像中,我们收集整理了二千五百幅包含防震锤的图像,建立了防震锤数据集,并手工标注防震锤Ground Truth,最后在此数据集中的检测图像上进行了防震锤检测,达到了 93%的准确率。(2)针对漏检防震锤情况,我们提出基于多视角匹配的方法对防震锤漏检情况进行补充,从而达到优化检测结果的目的。(3)基于深度学习的防震锤检测。针对本课题中无人机采集到的图像具有复杂场景,我们提出基于深度学习SSD模型的防震锤检测方法,并且设计了检测模型。最后对实验结果进行了分析。在我们构建的数据集上,SSD模型达到了 98%的准确率。
[Abstract]:In smart grid systems, intelligent inspection of transmission lines is an extremely important part. It is an important link to ensure the normal transportation of electric power. Therefore, the research on the detection of seismic hammer on the transmission line image shows the important value. The traditional inspection work of transmission line basically adopts the way of manual inspection. In recent years, unmanned aerial photography has been used by some power companies. However, the images captured by unmanned aerial vehicles have a wide field of view, complex background and perspective. The target resolution is not the same, which brings great challenge to the detection of earthquake hammer. Based on this background, this paper studies the detection and location algorithm of the shock hammer on transmission line by investigating the related work at home and abroad. In order to solve the problem of robust detection of anti-shock hammer in complex background and prepare for the fault diagnosis of anti-shock hammer, the main work of this paper is: 1) the detection of anti-shock hammer based on AdaBoost algorithm. The method based on template matching is only suitable for the detection of shock proof hammer with specific scene and angle, but not for the problem of this data set. This paper presents a method of detecting earthquake hammer based on AdaBoost algorithm. We used the Haar feature of the shock hammer and then trained it with the AdaBoost classifier. In the detection phase, because of the dependency between the hammer and the transmission line, we first extracted the transmission line and candidate the region of interest with the hammer. Finally, we compare and analyze the effects of different feature selection and different parameter settings on the experimental results. The experimental results are evaluated and analyzed from two aspects of accuracy and computational efficiency. In the field transmission line images collected by UAV, 2,500 images containing shock proof hammer are collected and arranged, and the seismic hammer data set is established. And manually annotate the shock hammer Ground Truth.Finally, the anti-shock hammer detection is carried out on the detection image of this data set, and the accuracy rate of 93% is reached. In order to optimize the detection results, we propose a method based on multi-angle matching to supplement the detection of earthquake proof hammer, which is based on depth learning. The images collected by UAV in this subject have complex scenes. We propose a method for detecting earthquake hammer based on depth learning SSD model, and design a detection model. Finally, the experimental results are analyzed. The accuracy of the model is 98% on our data set.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TM755

【参考文献】