第36卷第2期2023年2月传感技术学报CHINESEJOUＲNALOFSENSOＲSANDACTUATOＲSVol.36No.2Feb．2023项目来源:贵州省科学技术基金资助项目(黔科合基础［2016］1054);贵州省联合资金资助项目(黔科合LH字［2017］7226号);贵州大学2017年度学术新苗培养及创新探索专项项目(黔科合平台人才［2017］5788)收稿日期:2022－04－11修改日期:2022－05－26FusionTransformerandImprovedPANetforYOLOv5sTrafficSignDetection*ZHANGQian1，LIUZiyan1，2，3*，CHENYunlei1，WUYingyu1，ZHENGXuhui1(1．CollegeofBigDataandInformationEngineering，GuizhouUniversity，GuiyangGuizhou550025，China;2．StateKeyLaboratoryofPublicBigData，GuizhouUniversity，GuiyangGuizhou550025，China;3．InstituteofComputingTechnology，ChineseAcademyofSciences，Beijing100190，China)Abstract:Aimingatissuesoflowspeedoftrafficsigndetectionandextremeimbalanceoftargetsizeandcategory，aYOLOv5strafficsigndetectionalgorithmoffusedTransformerandimprovedPANetnetworkisproposed．Firstly，withoutincreasingmodelcomplexity，theendofbackbonenetworkisfusedwithTransformertoimprovenetworkfeatureextractioncapability．Secondly，thesmallsizeoftargetscaleintrafficsigndatasetscausespoordetectioneffectsofthescaledetectionlayerswithascaleof32timeslarger，thuscorrelationnetworklayersisnotused，andK-meansalgorithmisadoptedtoyieldappropriatepredictioncandidateframes．Thenthelossfunctionsisimprovedtoaddressextremeimbalanceproblemsbetweenpositiveandnegativesamples．Finally，theproposedalgorithmisappliedonJetsonAGXXavierplatformforvalidation．Experimentalresultsdemonstratethattheproposedalgorithmachievesbetterperformance，thedetectionprecisionandrecallrateimprove2．2%and0．7%，respectively，andthenumberofmodelparametersandcomputationalcomplexityreduce25．8%and10．1%respectivelyincomparisonwiththeoriginalnetwork．Thedetectionspeedof76FPSonXaviermeetstherequirementsforreal-timetrafficsigndetection，whichiseasilydeployedinrealscenarios．Keywords:trafficsigndetection;JetsonAGXXavier;Transformer;PANet;YOLOv5sEEACC:7230doi:10．3969/j．issn．1004－1699．2023．02．010融合Transformer和改进P...
