基于改进YOLOv7的高速铁路接触网小零部件定位算法研究  

Research on Localization Algorithm for Smal Components of High-speed Railway Catenary Based on Improved YOLOv7

作　　者：张雨婧 刘琦玉 郭维平 刘志刚[1,2] Zhang Yujing;Liu Qiyu;Guo Weiping;Liu Zhigang(Key Laboratory of Railway Industry on Smart Traction Power Supply,Southwest Jiaotong University,Sichuan,Chengdu 611756,China;School of Electrical Engineering,Southwest Jiaotong University,Sichuan,Chengdu 610031,China)

机构地区：[1]西南交通大学智能牵引供电铁路行业重点实验室,四川成都611756 [2]西南交通大学电气工程学院,四川成都610031

出　　处：《铁道技术标准（中英文）》2023年第8期7-16,共10页Railway Technical Standard(Chinese & English)

基　　金：国家自然科学基金(51977182)。

摘　　要：接触网支持装置零部件的正常工作状态对于高速铁路的安全运行至关重要,因此需要对零部件图片进行异常识别。检测车采集到的接触网全局图片中包含不同类别的零部件,为了高效、准确检测不同类别的零部件状态,需要从接触网全局图片中定位不同类别零部件。由于不同接触网零部件尺度差异较大,存在部分小尺度零部件特征较少、难以识别的问题,且目前检测方法主要针对一些大尺度零部件。本文根据现有的接触网零部件定位类别和缺陷类型,扩充目前的零部件定位数据,构建出有34个零部件种类的接触网多尺度零部件定位数据集。然后以YOLOv7定位模型为基础,根据接触网零部件定位的特点,改进网络模型,提高小零部件定位性能,并与常用的几种目标检测算法进行对比,验证本文定位模型的有效性。The proper functioning of the catenary support device components is crucial for the safe operation of high-speed railways.Therefore,it is necessary to identify anomalies according to the component images.The global images of the catenary collected by inspection vehicles,contain various categories of components.To efficiently and accurately detect the status of different categories of components,localization of these components is necessary from the global catenary images.Due to the significant scale differences among different catenary components,there are challenges in recognizing small-scale components with fewer features,and existing detection methods primarily focus on large-scale components.In this paper,we expanded the existing catenary component localization dataset to include 34 types of components at multiple scales s,considering the current localization categories and defect types.We then improved the YOLOv7 localization model based on the characteristics of catenary component localization to enhance the performance of localizing small-scale components.The effectiveness of the proposed localization model is validated through comparisons with several commonly used object detection algorithms.

关 键 词：接触网 YOLOv7 多尺度目标检测 深度学习 

分 类 号：U225.4[交通运输工程—道路与铁道工程]