摘要
针对当前列车轨道障碍物检测方法存在的小目标检测精度低、模型过大且部署成本高等问题,文中提出一种改进的YOLOv8-SGFE轨道侵限物检测模型。首先,为了减少网络的计算量,在小目标检测模块SPD-Conv的基础上,设计了一个SGConv模块,并用其替换YOLOv8主干层中的普通卷积层;其次,为了增强模型的感知能力,将高效多尺度注意力EMA与C2f-Faster模块相结合,构成C2f-Faster-EMA模块,并用其替换YOLOv8中的C2f模块;最后,将改进后的YOLOv8-SGFE模型应用于自制的铁路轨道侵限物数据集。与YOLOv8模型相比,文中模型参数量下降36.04%,FLOPs由28.7×10^(9)减少到19×10^(9),在模型计算量大幅降低的情况下,mAP提高2.5%。实验结果表明,所提算法具有更高的检测精度,模型参数量及计算负载更小,不仅适用于复杂环境下的轨道障碍物检测,同时更易于部署到移动端设备中。
An improved YOLOv8⁃SGFE track intrusion detection model is proposed to improve the detection accuracy for small objects,and reduce the large model size and the high deployment costs in the current train track obstacle detection methods.To reduce the network′s computational load,an SGConv module is designed based on the small object detection module SPD⁃Conv,and the SGConv module is used to replace the standard convolutional layers in the backbone of YOLOv8.To enhance the model′s perceptual capability,the efficient multi⁃scale attention(EMA)is combined with the C2f⁃Faster module to form the C2f⁃Faster⁃EMA module.The C2f⁃Faster⁃EMA module is used to replace the C2f module in YOLOv8.The improved YOLOv8⁃SGFE model is applied to a custom railway track intrusion dataset.In comparison with the YOLOv8 model,the proposed model′s parameters are decreased by 36.04%,and its FLOPs are reduced from 28.7×10^(9) to 19×10^(9).Despite the significant reduction in computational load,its mAP is increased by 2.5%.Experimental results demonstrate that the proposed algorithm achieves higher detection accuracy,with reduced model parameters and computational load,so it is suitable for detecting track obstacles in complex environments and easier to be deployed on mobile devices.
作者
冯庆胜
付明雨
姚泽圆
刘杨
梁天添
FENG Qingsheng;FU Mingyu;YAO Zeyuan;LIU Yang;LIANG Tiantian(School of Automation and Electrical Engineering,Dalian Jiaotong University,Dalian 116021,China;School of Computer and Communication Engineering,Dalian Jiaotong University,Dalian 116021,China)
出处
《现代电子技术》
北大核心
2025年第11期174-179,共6页
Modern Electronics Technique
基金
辽宁省教育厅科学研究项目(JYTMS20230008)
辽宁省教育厅基本科研项目(JYTMS20230037)。
