基于改进YOLOv3的铁路落石检测方法研究与实现

Research and Implementation of the Railway Rockfall Detection based on Modified YOLOv3

我国山区铁路沿线边坡崩塌落石侵入铁路限界将严重危及列车运行安全,当前铁路危岩落石的自动化监测系统存在误报、漏报及时效性低等缺点,各类监测系统对侵限落石的检测方法成为了制约系统性能指标的关键因素。本文收集了大量铁路真实场景下的崩塌落石样本,并选取Faster RCNN和YOLOv3算法进行了落石检测对比实验,实验表明YOLOv3算法与Faster RCNN算法检测精度相近,但检测速度更快,因此本文选取YOLOv3算法构建铁路边坡落石检测模型。由于YOLOv3算法对铁路落石群的检测准确度较低,本文采用多尺度检测以及调整网络超参数等方式对YOLOv3框架进行了改进,结果显示对落石群检测准确度提升效果较为明显,更能满足铁路沿线危岩落石检测的实际应用要求。

The slope collapse and rockfall intrusion into the railway clearance along the railway in mountainous areas of China seriously endangers the safety of train operation.At present,the automatic monitoring system of dangerous rockfalls on railways has the disadvantages of false alarms,missing reports,and low timeliness.The detection methods of various monitoring systems for rockfall intrusion have become the key factors restricting the system performance indicators.In this paper,a large number of collapse and rockfall samples of real railway scenes are collected,and the Faster RCNN and YOLOv3 algorithms are selected for rockfall detection comparison experiments.The experiments show that the accuracy of the YOLOv3 algorithm is similar to that of the Faster RCNN algorithm,but the detection speed is faster.Therefore,the YOLOv3 algorithm is selected to build a railway slope rockfall detection model in this paper.As the accuracy of the YOLOv3 algorithm for detecting rockfall clusters is low,this paper modified the YOLOv3 framework by using multi-scale detection and adjusting network hyperparameters.The results show that the accuracy of rockfall group detection is significantly improved,which can better meet the practical application requirements of dangerous rockfall detection along the railway.