基于深度学习的铁路限界快速检测算法

Fast detection algorithm of railway clearance based on deep learning

列车行驶环境快速、可靠、精准感知是列车安全、高效运行的前提和关键支撑技术。列车若无法提前感知非法侵入铁路限界范围内的异物,并在短时间内有效制动,将会导致严重的安全事故。为解决列车运行环境内影响运行安全的异物侵限问题,基于深度学习算法,提出一种铁路轨道限界快速检测算法。该方法首先采用预设行锚框的方式对图像进行划分,将传统分割算法的逐像素预测,转变为对每个行锚框进行逐网格预测,以达到显著降低算法计算量,并提高检测速度的目的。同时,通过快速识别图像中属于轨道部分的像素,结合轨道线的连续特征进行追踪,达到铁路轨道坐标的智能快速识别。最后根据标准轨距下限界的定义,对识别出的轨道线坐标和侵限区域进行扩充,以确定铁路异物入侵限界的范围。通过在真实轨道数据集上的对比实验,验证所提算法能以172FPS的速度快速检测铁路限界,且轨道线和限界区域的识别精度分别达到98.96%和98.12%,F1值分别达到99.68%和98.95%,限界区域检测的平均交并比MIoU达到96.88%,各类指标均取得当前最好的准确率和性能,满足高速列车对环境感知精度和速度要求,可以为异物侵限检测、目标跟踪和列车控制等环境感知及运行控制等下游任务提供基础,提升列车运行的安全性。

The fast,reliable,and accurate perception of driving environment is a fundamental and key technology for safe and efficient train running.The failure to detect foreign objectsthatintrude illegally into the railway clearance in advance and applyeffectivebrakes in a short time can lead to serious safety accidents.To address the problem of foreign object intrusion in the train running environment,a fast detection algorithm of railway clearance based on deep learning was proposed in this paper.First,the image segmentation algorithm based on deep learning was optimized by using row anchor boxes,which transforms the pixel-by-pixel prediction of traditional segmentation algorithms into a grid-by-grid prediction for each row anchor box.This optimization method significantly reduces the computational requirements of traditional segmentation algorithms and improves the detection speed.Second,the intelligent identification of the railway track coordinate position was achieved by promptly recognizing the pixel position of the track line in the image and combining it with the track line’s continuous characteristics.Finally,the scope of foreign body intrusion into the railway clearance was determined by expanding the coordinate position of the identified track line and intrusion area according to the definition of the railway clearance of standard gauge.Through comparative experiments,the proposed algorithm can detect railway clearance at 172 frames per second,with railway track and railway clearance recognition accuracies of 98.96%and 98.12%,respectively.The harmonic mean F1 scores of railway track and railway clearance detection are 99.68%and 98.95%,respectively,and the mean intersection over union(MIoU)detection is 96.88%.All these indicators achieved state-of-the-art accuracy and performance,which meets the environmental perception accuracy and speed requirements of high-speed trains.The proposed algorithm can provide a foundation for downstream tasks such as foreign body invasion detection,target tracking,train control,and environmental perception,as well as operation control,and improve the safety of train running.