基于LGF-Net的全天候轨道入侵异物智能检测系统  

All-weather intelligent detection system for railway intrusion obstacles based on LGF-Net

作　　者：赵宗扬 康杰虎[1] 梁健 叶涛[2] 吴斌[1] Zhao Zongyang;Kang Jiehu;Liang Jian;Ye Tao;Wu Bin(State Key Laboratory of Precision Measuring Technology and Instruments,Tianjin University,Tianjin 300072,China;School of Mechanical Electronic&Information Engineering,China University of Mining&Technology,Beijing 100083,China)

机构地区：[1]天津大学精密测试技术及仪器全国重点实验室,天津300072 [2]中国矿业大学(北京)机电与信息工程学院,北京100083

出　　处：《仪器仪表学报》2023年第9期285-299,共15页Chinese Journal of Scientific Instrument

基　　金：天津市交通运输科技发展计划项目(2022-09);北京市自然科学基金(L221018);光纤传感与系统北京实验室开放课题(GXKF2022001);天津大学自主创新基金(2023XHX-0019)项目资助。

摘　　要：针对轨道入侵异物为行车安全带来巨大威胁,而现有的轨道目标检测模型检测精度和速度难以平衡、复杂轨道环境中多尺度目标检测鲁棒性差等问题,提出了一种全天候高精度实时多尺度轨道入侵异物检测模型。该模型通过使用双分支结构和线性特征变换提升模型的特征提取速度;通过改进Transformer结构使轻量型模型能够建模全局上下文信息;通过设计高丰富度特征融合结构和轻量型注意力机制进一步提升模型的多尺度目标检测能力。此外,本文将该模型进行嵌入式移植并研制智能检测系统。实验结果表明,本文所提出的模型在实际轨道场景采集的数据集中检测精度和速度分别为94.93%和132 fps,比YOLOv5s高3.09%,能够满足在复杂轨道场景中高精度实时检测多尺度入侵异物的应用需求。Aiming at the enormous threat that railway intrusion obstacles pose to train operation safety,and the existing railway obstacle detection models have difficulty balancing detection accuracy and speed and poor multi-scale object detection robustness in complex railway environments,this article proposes an all-weather high-precision real-time multi-scale railway obstacle detection model.The model improves the feature extraction speed of the model by using dual-branch structure and linear operation.By modifying the Transformer structure,the lightweight model can model global contextual information.By designing high richness feature fusion structure and lightweight attention mechanism,the model′s multi-scale object detection ability is further improved.In addition,we embed the model and develop an intelligent detection system.The experimental results show that the proposed model has a detection accuracy and speed of 94.93%and 132 fps in the dataset collected from actual railway scenes,respectively,which is 3.09%higher than YOLOv5s.It can meet the application requirements of high-precision real-time detection of multi-scale obstacles in complex railway scenes.

关 键 词：轨道入侵异物 目标检测 深度学习 神经网络 检测系统 

分 类 号：U491.2[交通运输工程—交通运输规划与管理] TH39[交通运输工程—道路与铁道工程]