基于孪生多注意力网络的接触网侵限轻飘物跟踪  

Siamese Multi-attention Network-based Approach to Tracking of Light Object Intrusion into Overhead Contact System

作　　者：屈志坚[1,2] 张博语 朱琳 梁家敏 QU Zhijian;ZHANG Boyu;ZHU Lin;LIANG Jiamin(State Key Laboratory of Performance Monitoring and Protecting of Rail Transit Infrastructure,East China Jiaotong University,Nanchang 330013,China;School of Electrical and Automation Engineering,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学轨道交通基础设施性能监测与保障国家重点实验室,江西南昌330013 [2]华东交通大学电气与自动化工程学院,江西南昌330013

出　　处：《铁道学报》2024年第2期45-55,共11页Journal of the China Railway Society

基　　金：江西省高层次高技能领军人才培养工程(202223323);江西省自然科学基金(20232ACB204025);轨道交通基础设施性能监测与保障国家重点实验室开放课题(HJGZ2022203)。

摘　　要：针对接触网侵限轻飘物尺度变化大、遮挡干扰和铁路限界背景杂乱,易导致跟踪失败的问题,提出一种基于孪生多注意力网络的接触网轻飘物侵限跟踪新方法。引入三种注意力机制从更深层次提取轻飘物特征,通过空间注意力消除局部感受野限制,通过通道注意力突出轻飘物类别的通道特征,将交叉注意力聚焦于目标模板与搜索图像信息,采用空间正则化滤波器抑制背景干扰,融合各部分特征实现对侵限轻飘物的跟踪。采用OTB100数据集进行精度与准度实验,以国家重点实验室试验线采集的数据为算例进行实验,通过消融实验验证了新方法的有效性。结果表明:相比于相关滤波类SRDCF算法和深度学习类SiamRPN++、DaSiamRPN算法,新方法可获得更好的鲁棒性和准确性。A new method based on siamese multi-attention network was proposed to address the issues such as large scale variation of light objects intruding overhead contact system,occlusion interference and busy background of railroad clearance that may cause tracking failure.Three attention mechanisms were introduced to extract the flicker features from a deeper level,eliminate the local perceptual field restriction by spatial attention,highlight the channel features of flicker category by channel attention,focus the cross-attention on the contextual relationship between the target template and the search image,and suppress the background interference by spatial regularization filter before finally fusing the features of each part to achieve the tracking of the intruding flicker.Accuracy and accuracy experiments were conducted using the OTB100 dataset,and the data collected from the test line of the State Key Laboratory were used as arithmetic examples for experiments.The effectiveness of the new method was verified by ablation experiments.The results show that the new method can obtain better robustness and accuracy compared with the correlation filtering class SRDCF algorithm and the deep learning classes SiamRPN++and DaSiamRPN algorithms.

关 键 词：接触网限界 轻飘物 注意力机制 神经网络 空间正则化 

分 类 号：TM73[电气工程—电力系统及自动化] U225[交通运输工程—道路与铁道工程]