基于卷积神经网络的列车检测算法研究  

Train Detection Algorithm Based on Convolutional Neural Network

作　　者：祁颖 赵辉 张永鹏 王一力 宋亚京 QI Ying;ZHAO Hui;ZHANG Yongpeng;WANG Yili;SONG Yajing(Communication&Signaling Company,Beijing Subway Operation Co.,Ltd.,Beijing 100082,China;Subway Operation Technology Development Centre,Beijing Subway Operation Co.,Ltd.,Beijing 100044,China;Traffic Control Research Institute,Traffic Control Technology Co.,Ltd.,Beijing 100070,China)

机构地区：[1]北京市地铁运营有限公司通信信号分公司,北京100082 [2]北京市地铁运营有限公司地铁运营技术研发中心,北京100044 [3]交控科技股份有限公司交控研究院,北京100070

出　　处：《铁道运输与经济》2021年第3期74-78,87,共6页Railway Transport and Economy

基　　金：北京市科学技术委员会科技计划课题(Z191100002519005)。

摘　　要：列车检测作为列车自动驾驶的核心技术,可以有效地降低列车追尾等事故造成的人身危险和财产损失。为实现精准的列车检测,选用改进的卷积神经网络(PVANET)对输入图像进行特征提取,在此基础上,采用候选区域网络,从生成的特征图里滑动搜索,判断出图像中可能为列车的区域位置,并进一步采用快速区域卷积神经网络对每个候选区域进行分类,计算出其所属类别的置信度,同时精确定位列车。经验证,该方法适应范围广、鲁棒性高,可以有效地检测不同环境光强及不同朝向的列车,保障列车安全,为列车自动驾驶及辅助驾驶提供安全保障。As the key technique of the automatic driving, train detection can effectively reduce personal risks and the loss of property caused by train collision. To achieve precise train detection, the optimized convolutional neural network(PVANET) was adopted to extract the features from the input image. Then, the region proposal network was used for searching in the generated feature images, to determine the possible regional position of the train. The fast region convolutional neural network was adopted to classify each candidate region, calculate the confidence of its category, and accurately locate the train. The experimental results show that the proposed method with wide application range and high robustness, can effectively detect the light intensity and the trains in different directions, safeguard train safety, ensure the safety of trains and provide safety guarantees for automatic train driving and assisted driving.

关 键 词：卷积神经网络 候选区域网络 列车检测 智能交通系统 自动驾驶列车 

分 类 号：U298.1[交通运输工程—交通运输规划与管理]