616RadioCommunicationsTechnologyVol.49No.42023doi:10.3969/j.issn.1003-3114.2023.04.004引用格式:黄豪杰,唐宗璐,杨敏,等.基于改进YOLOv5的交通指示牌检测[J].无线电通信技术,2023,49(4):616-621.[HUANGHaojie,TANGZonglu,YANGMin,etal.ImprovedYOLOv5-basedTrafficSignDetection[J].RadioCommunicationsTech-nology,2023,49(4):616-621.]基于改进YOLOv5的交通指示牌检测黄豪杰1,唐宗璐2,杨敏3,李航4∗(1.南宁师范大学计算机与信息工程学院,广西南宁530100;2.北海市外国语实验学校,广西北海536000;3.合浦县实验学校,广西北海536100;4.广西民族大学人工智能学院,广西南宁530006)摘要:随着深度学习的不断发展,汽车自动驾驶已成为一种趋势,自动驾驶的安全问题是最重要的。其中,能准确识别复杂环境下密集的交通指示牌是保障安全驾驶的一个重要环节,针对目前检测模型对交通指示牌召回率不够高的问题,在YOLOv5的基础上提出了YOLOv5-ACB。经过300次的迭代训练,实验结果表明YOLOv5-ACB模型的mAP为62.9%、mAP50为83.6%、召回率为76.6%,相比原始的YOLOv5模型的mAP为62.45%、mAP50为82.6%、召回率为74.6%,均有较好的提升,说明所提出的改进模型降低了交通指示牌的错检和漏检率。关键词:YOLOv5;非对称卷积;TT100K;目标检测中图分类号:TP391文献标志码:A开放科学(资源服务)标识码(OSID):文章编号:1003-3114(2023)04-0616-06ImprovedYOLOv5-basedTrafficSignDetectionHUANGHaojie1,TANGZonglu2,YANGMin3,LIHang4∗(1.SchoolofComputerandInformationEngineering,NanningNormalUniversity,Nanning530100,China;2.BeihaiForeignLanguageExperimentalSchool,Beihai536000,China;3.HepuCountyExperimentalSchool,Beihai536100,China;4.SchoolofArtificialIntelligence,GuangxiUniversityforNationalities,Nanning530006,China)Abstract:Withthecontinuousdevelopmentofdeeplearning,autonomousdrivingofcarshasbecomeatrend.Andthesafetyofau-tonomousdrivingisundoubtedlythemostimportantissue,amongwhichtheabilitytoidentifydensetrafficsignsaccuratelyincomplexenvironmentiscriticaltoensuresafedriving.After300iterationsoftraining,experimentalresultsshowthattheproposedYOLOv5-ACBmodelhasanmAPof62.9%,anmAP50of83.6%andarecallrateof76.6%,whicharebetterthantheoriginalYOLOv5modelwithanmAPof62.45%,anmAP50of82.6%andarecallrateof74.6%.Thisindi...
