基于深度学习的铁轨图像识别技术  

Rail image recognition technology based on deep learning

作　　者：徐心慈 师秀霞 耿晨歌[1] 陈祥献[1] XU Xinci;SHI Xiuxia;GENG Chenge;CHEN Xiangxian(College of Biomedical Engineering&Instrument Science,Zhejiang University,Hangzhou 310027,China;UniTTEC Co.,Ltd.,Hangzhou 310000,China)

机构地区：[1]浙江大学生物医学工程与仪器科学学院,浙江杭州310027 [2]浙江众合科技股份有限公司,浙江杭州310000

出　　处：《铁道科学与工程学报》2024年第12期5232-5241,共10页Journal of Railway Science and Engineering

基　　金：浙江省科技计划项目(2022C01064)。

摘　　要：为了保证城轨交通中地铁的运营安全,避免铁轨轨道内障碍物的存在造成安全事故,有必要进行地铁铁轨识别。考虑到铁轨有细长连续的物理特性,同时受深度学习中的车道线检测任务启发,改进车道线检测网络CLRNet算法,提出针对铁轨图像识别的CLRNet-L算法。为解决铁轨轨道细长且难以被准确识别定位的问题,CLRNet-L使用特征金字塔网络进行高级特征与低级特征的提取与融合,通过自上而下的思想,首先用高层特征粗略地定位轨道,再利用与高层特征融合后的浅层特征进一步细化轨道从而实现铁轨识别。针对铁轨轨道线细长连续且因其颜色较深难以与周围环境区分的问题,在原CLRNet中引入注意力机制和多尺度聚合器,提出大核注意力模块捕获更多的上下文信息,加强对铁轨轨道的特征表示。由于在轨道交通领域中缺少针对铁轨识别任务的公开铁轨轨道数据集,使用在杭州5号线和6号线中采集的地铁运行时轨道照片制作轨道场景的数据集,数据集中包含直道、弯道和道岔场景的铁轨图像,该数据集用于进行铁轨识别实验,来验证算法的有效性。实验结果证明:CLRNet-L在自制数据集中达到了88.96%的MIoU和最快11.54 ms的检测速度,与其他算法相比有更高的精度和检测速度。研究结果能为地铁安全保障技术尤其是障碍物检测提供技术基础,保障地铁运营安全。In order to ensure the operation safety of the subway in urban rail transportation and avoid safety accidents caused by obstacles inside the railway tracks,it is necessary to conduct subway rail recognition.Considering that the railway tracks have slender and continuous physical characteristics,and inspired by the lane detection in deep learning,we proposed the CLRNet-L algorithm for railway track recognition,which is an improvement from the CLRNet algorithm.To solve the problem of long and thin railway tracks that are difficult to be accurately identified and localized,CLRNet-L used a feature pyramid network to extract and fuse high-level features and low-level features.Through the idea of top-down,the high-level features were first used to locate the railway tracks in a rough way,and then the shallow features fused with the high-level features were used to further refine the tracks so as to realize the identification of the railway tracks.In response to the problem of railway tracks that are difficult to distinguish from the surrounding environment due to their dark colors,attention mechanisms and multi-scale aggregators were introduced into the original CLRNet.We proposed a large kernel attention module to capture more contextual information and enhance the feature representation of railway tracks.Due to the lack of public railway track datasets for rail recognition in the field of rail transit,we used track photos collected during subway operation in Hangzhou Line 5 and Line 6 to create a dataset of track scenes.The dataset,which includes rail images of straight,curved,and turnout scenes,was used for rail recognition experiments to verify the effectiveness of the algorithm.Experimental results show that CLRNet-L achieved 88.96%MIoU and the fastest detection speed of 11.54 ms in the custom dataset,which has higher accuracy and detection speed compared with other detection algorithms.The research results provide a technical foundation for subway safety technology,especially obstacle detection,to ensure the safe

关 键 词：铁轨识别 图像处理 轨道交通 深度学习 计算机视觉 

分 类 号：U231.6[交通运输工程—道路与铁道工程] TP391[自动化与计算机技术—计算机应用技术]