Efficient HOG human detection

EfficientHOGhumandetectionYanweiPanga,YuanYuanb,n,XuelongLib,JingPancaSchoolofElectronicInformationEngineering,TianjinUniversity,Tianjin300072,P.R.ChinabCenterforOPTicalIMageryAnalysisandLearning(OPTIMAL),StateKeyLaboratoryofTransientOpticsandPhotonics,Xi’anInstituteofOpticsandPrecisionMechanics,ChineseAcademyofSciences,Xi’an710119,Shaanxi,P.R.ChinacSchoolofElectronicEngineering,TianjinUniversityofTechnologyandEducation,Tianjin300222,P.R.ChinaarticleinfoArticlehistory:Received21August2009Receivedinrevisedform11July2010Accepted23August2010Availableonline16September2010Keywords:ImageandvideoprocessingHumandetectionHOGFastalgorithmabstractWhileHistogramsofOrientedGradients(HOG)plusSupportVectorMachine(SVM)(HOG+SVM)isthemostsuccessfulhumandetectionalgorithm,itistime-consuming.Thispaperproposestwowaystodealwiththisproblem.OnewayistoreusethefeaturesinblockstoconstructtheHOGfeaturesforintersectingdetectionwindows.Anotherwayistoutilizesub-cellbasedinterpolationtoefficientlycomputetheHOGfeaturesforeachblock.Thecombinationofthetwowaysresultsinsignificantincreaseindetectinghumans—morethanfivetimesbetter.Toevaluatetheproposedmethod,wehaveestablishedatop-viewhumandatabase.Experimentalresultsonthetop-viewdatabaseandthewell-knownINRIAdatasethavedemonstratedtheeffectivenessandefficiencyoftheproposedmethod.&2010ElsevierB.V.Allrightsreserved.1.IntroductionObjectdetectionisanimportantstepofhigh-levelcomputervision.Reliableobjectdetectionisessentialtoimageunderstandingandvideoanalysis.Facesandhumanbodiesareamongthemostimportantobjectsinimagesandvideos.Therefore,facedetectionandhumandetectionhaveattractedconsiderableattentioninapplicationsofvideosurveillance[24,4],biometrics[25],smartrooms,drivingassistancesystems,socialsecurity[2],andeventanalysis.Inthispaper,wefocusonhumandetection.Detectinghumansinimagesischallengingduetothevariableappearance,illumination,andbackground[13,1,4].Generallyspeaking,detectinghumansinastaticimage(i.e.asingleimageoravideoframe)ismorechallengingthaninanimagesequence.Detectinghumansinanimagesequencecanusuallybeviewedasmovingobjectdetection.Somotioninformationcanbeused[5,6].Backgroundcanbemodeledandusedforforegrounddetection[3,6].However,inastaticimagethereisnomotionclueandthebackgroundcannotbemodeled.Usually,thecoreofdetectingobjectsinastaticimageiseffectivelymodelingtheintrinsiccharacteristicsoftheobjects.Thispaperlimitsitsscopetohumandetectioninstaticimage.Featureextractionandclassifierdesigningaretwokeystepsforreliablehumandetectioninastaticimage.Haar-likerectanglefeatureshavebeenusedfordetectinghumans[15]andfaces[16].Toimprovetherepresentativeanddiscriminatingcapacities,theoriginalHaar-likerectangleshavebeenextendedtorotatedfeatures[17],diagonalfeatures[17],andcenter-surroundedfeatures[14].TheHaar-likerectanglefeaturesencodetheintensitycontrastbetweenneighboringregions.Suchfeaturesaresuitableforfacedetection.Allfrontalfaceshavesimilarfacialcompo-nentsandthefacialcomponentshavefixedneighboringrelationship.Importantly,theintensitycontrastbetweenneighboringfacialpartsisrelativelystable.Butthesituationinhumanbodiesisquitedifferent.Theintensitycontrastbetweenregionsofahumanbodydependsontheappearanceofthehumanwear,whichvariesrandomly.SotheHaar-likefeatureisnotadiscriminatingfeatureforhumandetection.ThehumandetectionperformanceusingmerelyHaar-likerectanglefeaturesisfarfromacceptable.ContentslistsavailableatScienceDirectjournalhomepage::10.1016/j.sigpro.2010.08.010nCorrespondingauthor.E-mailaddress:[e-mail](Y.Yuan).SignalProcessing91(2011)773–781Scaleinvariancefeaturetransformation(SIFT)[19]isanalternativefeatureforhumandetection[20].PhungandBouzerdoum[21]proposedanovelfeaturecallededge-density(ED)forhumandetection.Anotherfeaturetakingadvantagesofedgeinformationisedgeorientationhisto-gram(EOH)[22].Regioncovariancematrix(RCM)[18,7]isamongthestate-of-the-artfeaturesforhumandetection.RCMisamatrixofcovarianceofsomeimagestatisticscomputedinsideanimageregion.Itisamatrix-formfeatureinsteadoftheusualvector-formfeature.Inthispaperweconcentrateonthemostsuccessfulandpopularvector-formfeature:histogramsoforientedgradients(HOG)[9,10,26].ItisinspiredbySIFTbutdifferentfromSIFT.HOGcanberegardedasadenseversionofSIFT.ItisshownthattheHOGfeaturesconcentrateonthecontrastofsilhouettecontoursagainstthebackground.Finally,itisnotedthatdifferenttypesoffeaturescanbecombinedtoenhancedetectionperformance[13].Forexample,Wangetal.[27]proposedtocombineHOGfeaturesandlocalbinarypattern(LBP)featuresinanelegantframework.Butfeaturefusion/combinationisbeyondthescopeofthepaper.Thesecondstepofhumandetectionisdesigningclassifier.Largegeneralizationabilityandlessclassifyingcomplexityaretwoimportantcriteriaforselectingclassifiers.Linearsupportvectormachine(SVM)[12]andAdaBoost[23]aretwowidely-usedclassifierssatisfy-ingthecriteria.Inthispaper,weplaceemphasesontheHOGfeatureandtheSVMclassifierandourcontributionslieinefficientlycomputingofHOGfeatur