摘 要:异物侵入受电弓对高速铁路运营安全危害极大,文章提出一种基于显著性和 YOLOv3 的受电弓异物侵限检测方法。首先,用 U2-Net 网络对采集到的图像进行显著性检测,准确定位受电弓区域;其次,将 YOLOv3 网络的预测尺度增加到 4 个,采用 K-means++ 算法重新计算先验框,用深度可分离卷积替换标准卷积的方法改进模型来提高准确度和速度;实验结果表明,改进后的 YOLOv3-R 模型检测准确度比 YOLOv3 提高了 8.68%,检测速度提高了 5.5%,能够快速有效地检测出受电弓上的异物。
关键词:异物检测;显著性检测;感兴趣区域;预测尺度;先验框初始化;深度可分离卷积
DOI:10.19850/j.cnki.2096-4706.2023.04.026
基金项目:中国铁路兰州局集团有限公司科技发展项目(LZJKY2022003-1)
中图分类号:TP391.4;TP18 文献标识码:A 文章编号:2096-4706(2023)04-0101-05
Detection of Pantograph Foreign Body Intrusion Based on Saliency and YOLOv3
WANG Tongli 1, GUO Youmin1, GAO Deyang2, FAN Yongqin2, HUANG Wenping2
(1.Institute of Mechanical and Electrical Technology, Lanzhou Jiaotong University, Lanzhou 730070, China; 2.China Railway Lanzhou Group Co., Ltd., Lanzhou 730070, China)
Abstract: Foreign body intrusion into pantograph has great harm to the operation safety of high-speed railway. This paper proposes a pantograph foreign body intrusion detection method based on saliency and YOLOv3. First, U2 -Net network is used to detect the saliency of the collected images and accurately locate the pantograph area. Secondly, the prediction scale of YOLOv3 network is increased to 4, the prior box is recalculated by K-means++ algorithm, and the standard convolution is replaced by depth separable convolution to improve the model to improve the accuracy and speed. The experimental results show, compared with YOLOv3, the detection accuracy of the improved YOLOv3-R model is increased by 8.68%, and the detection speed is increased by 5.5%, which can quickly and effectively detect foreign bodies on the pantograph.
Keywords: foreign body detection; saliency detection; region of interest; prediction scale; the prior box initialization; depth separable convolution
参考文献 :
[1] 韩志伟,刘志刚,张桂南,等 . 非接触式弓网图像检测技术研究综述 [J]. 铁道学报,2013,35(6):40-47.
[2] 赵晓娜,吴兴军,徐根厚 . 德国高速铁路接触网检测系统[J]. 中国铁路,2008(9):60-62.
[3] KUEN L K,HO S L,LEE T K Y,et al. A Novel IntelligentTrain Condition Monitoring System Coupling Laser Beam into Image Processing Algorithm [J].Transaction of the HongKong Institution of Engineers,2006,13(1):27-34.
[4] 董宏辉,葛大伟,秦勇,等 . 基于智能视频分析的铁路入侵检测技术研究 [J]. 中国铁道科学,2010,31(2):121-125.
[5] 李家才,陈治亚,王梦格 . 铁路入侵运动目标实时检测技术 [J]. 铁道科学与工程学报,2013,10(6):116-120.
[6] 段旺旺,唐鹏,金炜东,等 . 基于关键区域 HOG 特征的铁路接触网鸟巢检测 [J]. 中国铁路,2015(8):73-77.
[7] 王科理,高福来,杨鹏,等 . 基于深度学习的接触网鸟巢异物识别研究 [J]. 铁道机车车辆,2022,42(2):116-121.
[8] 王佳祺 . 基于卷积神经网络与稀疏编码的接触网关键部件及异物检测的研究 [D]. 成都:西南交通大学,2018.
[9] 郭晓静,隋昊达 . 改进 YOLOv3 在机场跑道异物目标检测中的应用 [J]. 计算机工程与应用,2021,57(8):249-255.
[10] 罗会兰,袁璞,童康 . 基于深度学习的显著性目标检测方法综述 [J]. 电子学报,2021,49(7):1417-1427.
[11] QIN X B,ZHANG Z C,HUANG C Y,et al. U2-Net: Going Deeper with Nested U-structure for Salient Object Detection [J/OL]. arXiv:2005.09007 [cs.CV].[2022-09-12].https://arxiv.org/abs/2005.09007.
[12] LIN T Y,DOLLAR P,GIRSHICK R,et al. Feature Pyramid Networks for Object Detection [C]//2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).Honolulu: IEEE,2017:936-944.
[13] HARTIGAN J A,WONG M A. Algorithm AS 136:A K-Means Clustering Algorithm [J].Journal of the Royal Statistical Society Series C Applied Statistics,1979,28(1):100-108.
[14] ARTHUR D,VASSILVITSKII S. K-Means++:The Advantages of Careful Seeding [C]//Proceedings of The Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms.New Orleans: Society for Industrial and Applied Mathematics,2007:1027–1035.
作者简介:王同丽(1997—),女,汉族,甘肃白银人,硕士研究生在读,研究方向:机器视觉、目标检测;郭佑民(1968—),男,汉族,甘肃陇西人,教授,研究方向:传感器与检测技术、智能检测;高德阳(1975—),男,汉族,甘肃兰州人,本科,研究方向:机车车辆;范永勤(1969—),男,汉族,甘肃兰州人,研究方向:机车车辆;黄文平(1980—),男,汉族,甘肃兰州人,研究方向:机车车辆。