搜索
354Vol.35,No.420094ACTAAUTOMATICASINICAApril,2009Adaboost1111.Adaboost\{.Adaboost,Haar.{.Haar,,.,,,Haar,AdaboostTP391ObjectDetectionbyCombinedModelBasedonCascadedAdaboostCUIXiao-Xiao1YAOAn-Bang1WANGGui-Jin1LINXing-Gang1AbstractSinglefeature-basedmodelalwaysmeetsthedi±cultiesofpoordetectionperformanceandslowdetectionspeedforobjectwithlargevariancesincolor,texture,andshape.AnovelcascadedandadditivemodelbasedoncascadedAdaboostclassi¯erisproposedinthispaper.ThiscombinedmodelconsistsoftwocascadedAdaboostclassi¯erswhichareindependentlytrainedwithedge-fragmentfeatureandHaarfeaturetodescribethewholeobjectandoneofitsstablecomponents,respectively.The¯nalclassi¯cationdecisionofthecombinedmodelismadeaccordingtothestageindexesbywhichasampleisrejectedoracceptedinthetwocascadedclassi¯ers.ExperimentsonseveraltestdatabasesshowthatthecombinedmodelcantakeadvantagesofthespeedmeritofHaarfeatureandtherobustnessofedge-fragmentfeature.Comparedwithsinglefeature-basedmodel,thedetectionperformanceofthecombinedmodelisgreatlyimproved.KeywordsObjectdetection,combinedmodel,edge-fragmentfeature,Haarfeature,cascadedAdaboost.Viola[1]HaarAdaboost.Haar,;Adaboost.,Haar.[2](),.[2],[3],.Shotton[4]Opelt[5],Chamfer,2008-01-212008-07-16ReceivedJanuary21,2008;inrevisedformJuly16,2008(60472028),(20040003015)SupportedbyNationalNaturalScienceFoundationofChina(60472028)andSpecializedResearchFundfortheDoctoralPro-gramofMinistryofEducationofChina(20040003015)1.1000841.DepartmentofElectronicEngineering,TsinghuaUniversity,Beijing100084DOI:10.3724/SP.J.1004.2009.00417Adaboost.,,,..,,.:,.,.[3](Principalcomponentanalysis,PCA)(Scale-invariantfeaturetransform,SIFT),:1)PCA-SIFT;2).Opelt[6],(Selected-features,SF)(Combined-model,CM).,Boost;,41835,.[6],,,[6].[7¡10].[11¡12].,,,.:,,,,;,,,.,AdaBoost.\{.,.Haar.Adaboost.,,Haar,.,,.11.1.1.1.1(Statisticalregionmerging,SRM)[13],.,..[5]..1.1.2,,,.Chamfer,DTE(xxx)=mineee2Efd(xxx;eee)g(1),E=feeeg,d.,,ChE;T(xxx)=1NTXttt2TDTE(ttt+xxx)(2),T=ftttg,NT;xxx..,.ST(E)=minjxxxjRfChE;T(xxx)g(3)DT(E)=d(pppopt;pppT)(4)pppopt=pppT+argminjxxxjRfChE;T(xxx)g(5),pppT;pppopt.,¸AdaboostCT(E)=S2T(E)+¸D2T(E)(6)1.1.3AdaboostAdaboost4:Adaboost419hT(E)=(aT;CT(E)thTbT;aTbT(7),aTbT.¸,¸.Adaboost,(1),,.1Fig.1Theglobalmodel1.2.,Haar.,,.1.31.3.1AdaboostAdaboost,NpN(oj!b)=po(Ro)N¡1(1¡Ro)po(Ro)N¡1(1¡Ro)+pb(Rb)N¡1(1¡Rb)=Ã1+pb(1¡Rb)po(1¡Ro)µpbpo¶N¡1!¡1(8),!b,o,popb,RoRbAdaboost.2(:pb=po=104,Rb=50%,Ro=99%).,.,,.,,Adaboost,,,,.2Fig.2Thesimulationresultoftherateoftrueobjectintherejectedsamplesvs.thestageindex1.3.2,Adaboost\{.Haar,,,,.,.,,,.,,,.3().3,Rnn,Pnn..,,.,NcNg,420353{Fig.3Frameworkofthe\cascaded-additivealgorithmpN(oj!o)=Ã1+pb(1¡Rcb)po(1¡Rco)µRcbRco¶Nc¡1µRgbRgo¶Ng!¡1(9),!o,RcoRcb,RgoRgb.(9),.,,Adaboost.,,:1),NN=Nc+Ng(10),Nc,Ng.2),NAlast=NRlast¡logµµRcbRco¶(1¡Rco)(1¡Rcb)¶logµRgbRgo¶(11),NAlast,NRlast.(11):,Ã1+pb(1¡Rcb)po(1¡Rco)µRcbRco¶Nc¡1µRgbRgo¶NRlast!¡1=Ã1+pbpoµRcbRco¶NcµRgbRgo¶NAlast!¡1µRgbRgo¶NRlast¡NAlast=µRcbRco¶(1¡Rco)(1¡Rcb)NAlast=NRlast¡logµµRcbRco¶(1¡Rco)(1¡Rcb)¶logµRgbRgo¶1.3.3,DRcm=n(Ro)Nc¡1(1¡Ro)(Ro)Ng+(Ro)Nc(Ro)Ng¡m=nRN¡1o(1¡Ro)+RN¡mo(12)FPRcm=n(Rb)Nc¡1(1¡Rb)(Rb)Ng+(Rb)Nc(Rb)Ng¡m=nRN¡1b(1¡Rb)+RN¡mb(13),Ro,Rb,n{,AdaboostDRsm=RNo(14)FPRsm=RNb(15)(12)»(15),4()ROC(Receiveroperatingcharacteristics),:Ro=99%,Rb=50%.,Adaboost.5().n,.,,,.n=0,,.4:Adaboost4214{Fig.4Simulationresultofcomparisonbetweencascaded-additivealgorithmandclassi¯erofsinglefeature5Fig.5Detectionperformancewithdi®erentbranchnumbers1.3.4,,,6.,.,,.,,,,,,,.2,UIUC.,6Fig.6Geometricmodel:Adaboost100%50%.10100,10.1.05.,50%.2.1..,200,75£50.,15643000,24£24.50Adaboost,19Adaboost.(2:5;2:5;4:5;2).46.,,1.055,.292,320..N(50»61)ROC.{n7.,(95%),(0.33/).1419.1413.2.1.1742235.4.7{Fig.7Experimentalresultofcomparisonbetweencascaded-additivealgorithmandclassi¯erofsinglefeature2.1.2,13,.13,ROC.8,{.1.3.3,{\.85.8{Fig.8Experimentalresultofcomparisonbetweencascaded-additivealgorithmandsimplecascadedalgorithm2.1.3,..,Haar,.,,O(nlogn),.,,.39N54.2.1.1,,,950£850950£850.,950£850,222.,,17.8%.,80%,,,.,.2.1.49,,.,,,.9Fig.9Samplesofdogdetectionresults4:Adaboost4232.2UIUC2.2.1UIUC,,.18862000,24£24;,550,40£100.,.[4],.,16,7.4.170,200([4]164,193).,.,,,(1;3;1;0)(3;1;1;0),25/65/3.,1.053.,,,.1,.,{,.1Table1ComparisonwithglobalmodelandcomponentmodelR-PEER93%88%{96%2.2.2[4,14¡16].[14,16],.[15]PCA.[4],,[4].,,25,.40£100.,(10).10UIUCFig.10SamplesofUIUCside-viewcardetectionresults2[14¡16][4].R-PEER,{.2Table2Comparisonwithotherstate-of-the-artalgorithmsR-PEERAgarwal[14]79%Fergus[15]88.5%Leibe[16]91.0%Shotton[4]92.8%{96%3\{,Adaboost.Haar,.Haar,..,.,{Adaboost.,.42435References1ViolaP,JonesMJ.Rapidobjectdetectionusingaboostedcascadeofsimplefeatures.In:Proceedingsofthe2001IEEEComputerSocietyConferenceonComputerVisionandPat-ternRecognition.Hawaii,USA:IEEE,2001.511¡5182LeibeB,SchieleB.Analyzingappearanceandcontourbasedmethodsforobjectcategorization.In:ProceedingsofIEEEComputerSocietyConferenceonComputerVisionandPat-ternRecognition.Wisconsin,USA:IEEE,2003.409¡4153ZhangW,YuB,ZelinskyGJ,SamarasD.Objectclassrecognitionusingmultiplelayerboo