基于YOLO v4的铁道侵限障碍物检测方法研究  

Research on detection method of railway intrusion obstacles based on Yolo v4

作　　者：刘力 苟军年[1] LIU Li;GOU Junnian(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,甘肃兰州730070

出　　处：《铁道科学与工程学报》2022年第2期528-536,共9页Journal of Railway Science and Engineering

基　　金：光电技术与智能控制教育部重点实验室开放课题(KFKT2018-14)。

摘　　要：铁路侵限异物的自动检测是未来实现铁路智能化的重要组成部分。由于随机的侵限行为可能导致严重的行车后果,研究可以实现连续检测列车运行前方区域状况的技术,是保障列车出行安全的现实需求。针对传统侵限异物检测方法检测类别单一和时效性差的不足,提出一种基于YOLO v4检测网络的侵限异物检测模型。在锚框(anchor)的选择上,通过对K-means算法聚类中心的选取方法进行改进,用欧式距离度量替换随机选择的方法,从而获得更具代表的anchor尺寸;在YOLO v4网络的基础上,通过在骨干网络和特征融合网络之间加入压缩和激励模块,在不增加检测时间的同时提升了检测效果;在侵限检测模型的训练方面,使用公共数据集和自制异物侵限数据联合训练的方式提高了模型的泛化能力。在侵限异物测试集上对训练好的模型进行测试,结果表明:该方法对常见异物的平均检测精度达到90.2%,检测速度为53 fps,与Faster R-CNN相比检测精度相差较少的情况下,检测精度有大幅提升。改进的检测模型达到了预期设计目标,可以为铁路侵限异物检测智能化的研究提供参考。The automatic detection of foreign objectsinvading railroad limitsis an important part of the future realization of railroad intelligence.The random violation behavior may lead to serious consequences of train operations.Therefore,studying the technology that can continuously detect the condition of the areasin front of running train is the practical demand to ensuretrain operation safety.By aiming at the shortcomings of thesingle detection category and poor timeliness of traditional methods for detecting invasive foreign objects,a new detection model of invading foreign objectswas proposed based on the YOLO v4 detection network.In terms of selecting anchors,the method of Euclidean distance measurement was used toreplace the method of random selection,and the selection method of K-means clustering center was improved to obtain more representative anchor size by clustering.Based on the YOLO v4 network,by adding Squeeze-and-Excitation modules between the backbone network and the feature fusion network,the detection effect was improved without increasing the detection time.In terms of training the invasiondetection model,the generalization ability of the model was improved by combining public and customizeddata sets of foreign objects invading limits.After testing the trained model against the data sets of invading foreign objects,it shows that the average detection accuracy of this method is 90.2%and the detection speed is 53 fps.Compared with the results by the Faster R-CNN,the detection accuracy is greatly improved when there is less difference in detection accuracy.The improved detection model achieves the expected design goal and can provide a reference for intelligent detections of foreign objects invading railroad limits.

关 键 词：侵限异物 目标检测 YOLO v4 聚类 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]