基于改进型SSD算法的铁路货场异物侵限小目标检测研究  

Research on Small Object Detection of Foreign Object Intrusion in Railway Freight Yard Based on the Improved SSD Algorithm

作　　者：李建国 陈敬涛[3] 张伟 李斌 LI Jianguo;CHEN Jingtao;ZHANG Wei;LI Bin

机构地区：[1]中国铁路兰州局集团有限公司货运部 [2]兰州交通大学高原铁路运输智慧管控铁路行业重点实验室,兰州730070 [3]北京佳讯飞鸿电气股份有限公司,北京100089

出　　处：《铁道通信信号》2024年第7期57-62,共6页Railway Signalling & Communication

基　　金：中国铁路兰州局集团有限公司科技开发计划(LZJKY2022060-2,LZJKY2023006-1)。

摘　　要：为解决铁路货场异物侵限场景中小目标检测难度大、准确率不高、检测效果不佳的问题,提出基于特征金字塔网络的改进型单阶多框检测(SSD)算法。通过研究现阶段铁路货场业务管理现状、异物侵限场景及对应的检测技术,对小目标检测现存问题进行归类总结;通过在SSD算法的检测网络部分增加不同特征层信息的金字塔网络结构,提高小目标检测效率。根据改进前后2种算法在铁路货场异物侵限场景的试验数据对比,得出改进型SSD算法推理阶段的检出精度更高,可有效提高小目标检测效率和准确率,为铁路货场智能化安全管控提供有力的技术支撑。To solve the problems of high difficulty,low accuracy and poor effects of small object detection in the scene of foreign object intrusion in railway freight yards,an improved single shot multibox detection(SSD)algorithm based on Feature Pyramid Networks is proposed.The existing problems of small object detection are classified and summarized by studying the current business management status of railway freight yards,foreign object intrusion scenes,and corresponding detection technologies.A pyramid network structure with different feature layer information is added to the detection network section of the SSD algorithm to improve the efficiency of small object detection.As indicated by the test data of two algorithms,i.e.,improved/unimproved SSD algorithms,in the scene of foreign object intrusion in railway freight yards,the improved SSD algorithm has better detection accuracy in the inference stage,and thus can effectively improve the efficiency and accuracy of small object detection and provide strong technical support for intelligent safety management and control of railway freight yards.

关 键 词：深度学习 小目标检测算法 铁路货场 异物侵限 SSD算法 特征金字塔网络 

分 类 号：U285.4[交通运输工程—交通信息工程及控制] TN87[交通运输工程—道路与铁道工程]