基于改进YOLOv5的降雪天气高速列车障碍物检测  

Obstacle Detection of High-Speed Train in Snowfall Weather Based on Improved YOLOv5

作　　者：马晓君[1] 王栋 刘德胜[2] 梁晨 MA Xiao-jun;WANG Dong;LIU De-sheng;LIANG Chen(School of Mechanical Engineering,Jiamusi University,Jiamusi Heilongjiang 154007,China;School of Information and Electronic Technology,Jiamusi University,Jiamusi Heilongjiang 154007,China)

机构地区：[1]佳木斯大学机械工程学院,黑龙江佳木斯154007 [2]佳木斯大学信息电子技术学院,黑龙江佳木斯154007

出　　处：《计算机仿真》2025年第1期155-161,451,共8页Computer Simulation

基　　金：黑龙江省优秀科技创新团队项目(2019-KYYWF-1335);黑龙江省教育厅基本科研业务费基础研究项目(2020-KYYWF-0261)。

摘　　要：针对降雪天气造成的铁路场景不清晰,以及遮挡造成的目标误检率等问题,提出了一种基于改进的YOLOv5的铁路障碍物入侵检测网络模型。在原有算法基础上引入坐标注意力检测机制,提高特征的提取能力,增强对遮挡目标及小目标的检测能力;提出Focal-SIoU边界框回归损失函数,加快训练的收敛速度并提升预测框的定位精度;引入RepGFPN提高网络的检测速度,保证识别的实时性。在数据集RD和VOC 2012上的实验结果表明,提出的算法与原YOLOv5算法相比,mAP_(@0.5)分别提高了6.1%和2%,检测速度分别达到64FPS和67FPS,表明提出的算法可以在降雪的天气下快速、准确地检测出障碍物。Aiming at the problems of unclear railway scene caused by snowfall weather and target false detection rate caused by occlusion,a railway obstacle intrusion detection network model based on improved YOLOv5 network is proposed.Based on the original YOLOv5 algorithm,this paper introduces the coordinate attention detection mechanism to improve the feature extraction ability and enhance the detection ability of occluded targets and small targets.The Focal-SloU bounding box regression loss function is proposed to accelerate the convergence speed of training and improve the positioning accuracy of the prediction box.RepGFPN is introduced to improve the detection speed of the network and ensure the real-time recognition.The experimental results on the data sets RD and VOC 2012 show that compared with the original YOLOv5 algorithm,the mAP@0.5 is increased by 6.1%and 2%respectively,and the detection speed is 64 FPS and 67 FPS respectively.It shows that the proposed algorithm can quickly and accurately detect obstacles in snowy weather.

关 键 词：复杂天气 障碍物识别 高速列车 神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]