基金项目:国家自然科学基金项目(61563005)收稿日期:2021-06-02修回日期:2021-06-10第40卷第4期计算机仿真2023年4月文章编号:1006-9348(2023)04-0128-06基于YOLOv4的车辆与行人检测网络设计谭光兴,岑满伟∗,苏荣键(广西科技大学电气与信息工程学院,广西柳州545616)摘要:针对YOLOv4网络模型参数量大,难以在资源有限的设备平台上运行的问题,提出一种对YOLOv4轻量化的车辆和行人检测网络。以MobileNetV1为主干网络,将PANet和YOLOHead结构中的标准卷积替换成深度可分离卷积,减少模型参数量;同时利用跨深度卷积结合不同膨胀率的空洞卷积构建特征增强模块,改善不同预测层对车辆和行人尺度变化的适应能力,提高网络的检测精度。实验结果表明,上述网络模型大小为45.28MB,检测速度为44FPS,相比YOLOv4模型大小减少81.44%,检测速度提升91.30%,在PASCALVOC2007测试集上,检测精度达到86.32%,相比MobileNetV1-YOLOv4原网络提高1.29%的精确度,能够满足实时高效的检测要求。关键词:深度学习;目标检测;特征增强;轻量化中图分类号:TP183文献标识码:BDesignofVehicleandPedestrianDetectionNetworkBasedonYOLOv4TANGuang-xing,CENMan-wei∗,SURong-jian(SchoolofElectricalandInformationEngineering,GuangxiUniversityofScienceandTechnology,LiuzhouGuangxi545616,China)ABSTRACT:AimingattheproblemthatthenumberofYOLOv4networkmodelparametersislargeandisdifficulttorunondeviceplatformswithlimitedresources,alightweightvehicleandpedestriandetectionnetworkforYOLOv4isproposed.UsingMobileNetV1asthebackbonenetwork,thestandardconvolutioninthePANetandYOLOHeadstructureisreplacedwithadeepseparableconvolutiontoreducetheamountofmodelparameters;atthesametime,thecross-depthconvolutioncombinedwiththeatrousconvolutionwithdifferentdilationratesisusedtoconstructafeatureenhancementmodule,improvingtheadaptabilityofdifferentpredictionlayerstothescalechangesofvehiclesandpedestriansandincreasingthedetectionaccuracyofthenetwork.Theexperimentalresultsshowthatthesizeofthenetworkmodelis45.28MB,whichis81.44%smallerthanthatoftheYOLOv4model;itsdetectionspeedis44FPS,anincreaseof91.30%comparedtotheYOLOv4model.OnthePASCALVOC2007dataset,thedetectionaccuracyreaches86.32%,anincreaseof1.29%comparedtotheoriginalMobileNetV...
