高速铁路无砟轨道车载检测图像异物识别方法  

Onboard Image Recognition Method for Detection of Foreign Objects on High Speed Railway Ballastless Tracks

作　　者：杨怀志 刘洪润 宋浩然 顾子晨 王浩然 王乐 杜馨瑜 戴鹏 YANG Huaizhi;LIU Hongrun;SONG Haoran;GU Zichen;WANG Haoran;WANG Le;DU Xinyu;DAI Peng(Beijing-Fuzhou Passenger Dedicated Railway Line Anhui Co.,Ltd.,Hefei Anhui 230031,China;Beijing-Shanghai High Speed Railway Co.,Ltd.,Beijing 100089,China;Infrastructure Inspection Research Institute,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China;Beijing IMAP Technology Co.,Ltd.Beijing 100081,China;China Academy of Railway Sciences,Beijing 100081,China)

机构地区：[1]京福铁路客运专线安徽有限责任公司,安徽合肥230031 [2]京沪高速铁路股份有限公司,北京100089 [3]中国铁道科学研究院集团有限公司基础设施检测研究所,北京100081 [4]北京铁科英迈技术有限公司,北京100081 [5]中国铁道科学研究院,北京100081

出　　处：《中国铁路》2024年第4期8-14,共7页China Railway

基　　金：国家自然科学基金面上项目(52272427)。

摘　　要：针对高速铁路实际运营中易出现的无砟轨道异物问题,提出一种高速铁路无砟轨道异物图像识别方法。该方法基于改进的DeepLab无砟轨道异物语义分割模型,利用该模型对轨道图像的分割结果,可准确获取异物的像素级信息。为提高异物检出率和精确率,在模型的主干网络中引入通道注意力机制,用于关联图像上下文信息,实现模型对待识别区域的加权约束。在此基础上,针对无砟轨道异常检测中样本类别分布不平衡影响模型的问题,对模型的损失函数进行类别分配占比均衡的改进。试验结果表明,该方法可在像素级别上实现对于多种类型无砟轨道异物的识别,在测试集上检测精确率达到90%,检出率保持在95%以上。An image recognition method was developed to detect foreign objects on ballastless tracks of high speed railways,thus addressing the recurrent issue of foreign object presence on these tracks during operations.This method is based on the improved DeepLab semantic segmentation model of foreign objects on ballastless tracks.The pixel-level information of foreign objects can be accurately captured by using the track image segmentation results from this model.To improve the foreign object detection rate and accuracy,a channel attention mechanism was introduced into the backbone network of the model to correlate image context information and enable weighted constraints of the model on areas to be detected.Furthermore,the loss function of the model was improved by balancing the category allocation proportion to address the unbalanced distribution of sample categories affecting the model in anomaly detection of ballastless tracks.The test results demonstrate that this method can detect various types of foreign objects on ballastless tracks at the pixel level,achieving a detection accuracy of 90%and a detection rate above 95%on the test set.

关 键 词：高速铁路 无砟轨道 轨道异物 图像识别 异常检测 语义分割 注意力机制 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]