基于孪生网络的高铁线路灾害检测数据增广方法

Data augmentation method of high-speed railway line disaster detection based on Siamese network

在铁路运输过程中除了以人员入侵为代表的人为灾害,以落石、泥石流为代表的自然灾害对列车的运行的安全同样构成了极大的威胁。由于落石滑坡等的突发性,仅靠常规的人工检测,很难快速、有效、准确地进行监测,因此,针对高速轨道的数据收集困难,样本量较少等问题,结合孪生网络在变化检测中特征提取能力强等优点,提出了基于孪生网络和轨道落石检测的数据增广方法,该数据增广方法包括随机抖动和随机膨胀,得到的样本数据利用孪生网络的双输入,构建正负样本。负样本的选取首先是利用前景图片构建负样本,其次是使用帧差法热图扩充负样本,解决负样本难收集的问题。实验结果表明,加入本数据增广方法后线路灾害检测算法的识别准确率显著提升,且不影响检测速度,并且对于未经充分标注训练目标也能做出较为准确的检测,有效解决了线路灾害实时检测算法样本不足的问题。

In the process of railway transportation,in addition to man-made disasters represented by personnel invasion,natural disasters represented by rockfall and debrisflow also pose a great threat to the safety of train operation.Due to the suddenness of falling rocks and landslides,it is difficult for the railway sector to monitor quickly and accurately only by people checking and patrolling the railroad tracks.Because it is very difficult to collect high-speed track data,and the amount of data is small,and because of the advantages of Siamese network in feature extraction in change detection,we propose a data augmentation method based on Siamese network and rockfall detection.The data augmentation method includes random jitter and random expansion,and then we use the obtained data to construct positive and negative samples,and finally input them into the Siamese network for training.In the selection of negative samples,the foreground image is used to construct negative samples first,and then the frame difference method is used to construct heat maps to expand negative samples to solve the problem of difficult collection of negative samples.The experimental results show that the recognition accuracy of the line disaster detection algorithm is significantly improved after adding the data augmentation method,and the detection speed is not affected,and the new data can also be accurately detected,which effectively solves the problem of insufficient samples of high-speed railway line disaster real-time detection algorithm.