基于FasterNet和YOLOv8s改进的铁路异物入侵快速检测方法  

Improved rapid detection method for foreign object intrusion on railroads based on FasterNet and YOLOv8s

作　　者：姜香菊[1] 冯海照 李涛 JIANG Xiangju;FENG Haizhao;LI Tao(School of Automation and Electrical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;School of Mechanical Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)

机构地区：[1]兰州交通大学自动化与电气工程学院,兰州730070 [2]兰州交通大学机电工程学院,兰州730070

出　　处：《北京交通大学学报》2024年第5期39-48,共10页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：甘肃省科技计划项目(23JRRA868)。

摘　　要：铁路轨道异物入侵对列车行车安全构成潜在威胁,情况紧急时可能导致列车脱轨倾覆和人员伤亡.针对现有铁路异物入侵检测模型在边缘设备上无法保证实时检测的问题,提出一种基于Faster⁃Net和YOLOv8s改进的铁路异物识别算法.首先,采用参数量更小的FasterNet网络替代YOLOv8s的CSPDarkNet53主干网络进行特征提取,以减小模型参数量和计算量.然后,结合FasterNet中部分卷积思想,设计FasterBlock模块替代YOLOv8s颈部的C2f模块,实现多尺度特征融合,从而进一步减小模型参数量.最后,为解决网络轻量化导致的模型检测精度下降问题,重新设计BiFPN-A特征融合结构,采用Fusion替代Concat操作进行张量拼接操作,通过FasterBlock模块和Fusion实现跳跃连接的特征图融合,并在每一层FasterBlock模块之前引入无参注意力机制SimAM,保证改进后的整体模型在实现轻量化的同时能够有效防止检测精度的大幅下降.结果表明:在精度仅损失0.2%的情况下,改进后的模型尺寸减小60.89%,模型参数降低61.8%,计算量减小45.1%.Railway track obstructions pose potential threats to train operation safety,with severe inci-dents possibly leading to derailments,overturns,and casualties.To address the challenge of achieving real-time detection on edge devices,where existing railway intrusion detection models often fail,this paper proposes an improved railway obstruction detection algorithm based on FasterNet and YOLOv8s.First,FasterNet,a network with fewer parameters,replaces the CSPDarkNet53 back-bone of YOLOv8s for feature extraction,reducing both the parameters and computational complexity.Second,inspired by partial convolution in FasterNet,a FasterBlock module is introduced to replace the C2f module in YOLOv8s’neck,enabling multi-scale feature fusion and further decreasing model parameters.Finally,to mitigate accuracy loss caused by model lightweighting,a redesigned BiFPN-A feature fusion structure is proposed.In this structure,Fusion operations replace Concat for tensor concatenation,achieving feature map fusion via FasterBlock and Fusion.Additionally,a parameter-free attention mechanism SimAM is integrated before each FasterBlock,ensuring that the lightweight model maintains robust detection accuracy.The results demonstrate that the improved model achieves a 60.89%reduction in size,a 61.8%decrease in parameters,and a 45.1%reduction in computational complexity,with only a 0.2%loss in detection accuracy.

关 键 词：目标检测 YOLOv8 轻量化 SimAM 铁路异物 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]