基于深度学习的铁路异物侵限快速检测方法  

Fast detection method of railway foreign object intrusion based on deep learning

作　　者：王辉[1,2,3] 姜朱丰 吴雨杰 范自柱 罗国亮 杨辉 WANG Hui;JIANG Zhufeng;WU Yujie;FAN Zizhu;LUO Guoliang;YANG Hui(School of Software,East China Jiaotong University,Nanchang 330013,China;Jiangxi Provincial Key Laboratory of Advanced Control and Optimization,Nanchang 330013,China;Nanchang Virtual Reality Research Institute Co.,Ltd.,Nanchang 330000,China)

机构地区：[1]华东交通大学软件学院,江西南昌330013 [2]江西省先进控制与优化重点实验室,江西南昌330013 [3]南昌虚拟现实研究院股份有限公司,江西南昌330000

出　　处：《铁道科学与工程学报》2024年第5期2086-2098,共13页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(61991401,U2034211);江西省自然科学基金资助项目(20224BAB212014,20232ABC03A04);教育部人文社会科学研究项目(22YJCZH168)。

摘　　要：针对列车运行环境内因意外突发事件所造成的异物侵限而影响列车安全运行的问题,在被广泛应用于工业领域的YOLOv3目标检测模型的基础之上,提出一种融合轨道限界和侵限异物识别的快速检测方法。首先,以ResNet-18网络作为铁路限界检测的主干网络,利用辅助检测模块提升限界检测精度,达到特征提取速度快,语义信息丰富充足等目标。同时采用基于行锚框的分割算法检测轨道线坐标位置,结合标准轨距下的限界定义确定铁路异物入侵限界的范围,以减少侵限异物检测的区域。其次,设计基于Octave卷积的层内多尺度残差模块,将单通道卷积变为基于图像频率的双通道卷积,以降低卷积计算量,进一步提升异物侵限算法的检测速度。最后,引入空间金字塔模块和特征自适应融合模块,实现高、低级语义信息交换,从而增加网络对不同尺度目标的感知能力,并减少语义冲突问题。通过对比实验验证异物侵限检测算法的精度、速度和有效性。实验结果表明,所述方法能以172帧/s的速度对轨道位置和限界区域进行检测,精确度达98.12%。与其他算法相比,在大中小3种目标尺度上都超越了其他对比算法。所提出的融合轨道限界和侵限异物检测的方法,在保证精度的前提下,速度达到YOLOv3算法的2倍,能够满足列车对侵限异物的实时检测需求。The intrusion of foreign objects poses a great risk to the safety of the train and its passengers,and these objects can easily cause unexpected accidents.To address this problem,this paper proposed a rapid detection method by integrating track boundary detection with intrusion foreign object detection,employing the popular YOLOv3 model that was widely used in the industrial field.Firstly,the ResNet-18 network was used as the backbone,and an auxiliary detection module was added to improve detection accuracy.This approach could achieve fast feature extraction and capture sufficient semantic information.Simultaneously,a segmentation algorithm based on row anchor frames was used to detect the coordinates of the track lines.This article combined it with the definition of the railway foreign object invasion limit under standard gauge to reduce the detection area for intrusion objects.Secondly,an intra-layer multi-scale residual module based on Octave convolution was designed to convert the single-channel convolution into dual-channel convolution,significantly reducing the computational cost of convolution and further improving the detection speed.Finally,a spatial pyramid module and a feature adaptive fusion module were introduced to facilitate the exchange of high-level and low-level semantic information,thereby enhancing the network's ability to perceive targets at different scales and reducing semantic conflicts.The accuracy,speed,and effectiveness of the foreign object intrusion detection algorithm are verified through comparative experiments.The experimental results demonstrate that the proposed method achieves a detection speed of 172 frames per second,with an accuracy of 98.12%.Compared with other algorithms,the proposed method can achieve twice the speed of the YOLOv3 algorithm while maintaining accuracy.It also surpasses other comparison algorithms in detecting targets at large,medium,and small scales.The method can meet the real-time detection requirements for intrusion objects,and provide an efficient and ac

关 键 词：异物侵限检测 Octave卷积 行锚框 铁路限界检测 空间金字塔 特征自适应融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] U216.3[自动化与计算机技术—计算机科学与技术]