发表在CVPR13上的行人检测方法
Thispapermainlyfocusesonpairwisepedestriansforsimplicity.WhenthereareN(>2)pedestriansinalocalimageregion,pair-wiserelationshipisstillabletorepresenttheirvisibilityrelationships.Meanwhile,ourapproachcanbeextendedforconsideringN(>2)windowssimultane-ously.DenotethefeaturesofNwindowsbyx,denotethelabelforthenthwindowbyyin(1)isextendedto:
n∈{0,1},thep(y1|x1,x2)p(y1|x)=
Np(y
p(y 1,y2,···,yN|x)=1,yn=u).
y2,···,yN
u
n=2
(11)
When N
n=2yn=u,amutualvisibilitydeepmodelisconstructedforupedestrians,similartothemutualvisibil-itydeepmodelforpair-wisepedestrians.
5.ExperimentalResults
Theproposedframeworkisevaluatedonfourpubliclyavailabledatasets:Caltech-Train,Caltech-Test[9],ETH[13]andPETS20092.Inourimplementation,theDPMin[14]withthemodi edHOGfeaturein[14]isusedforpartdetectionscores.Thedeformationamongpartsarear-rangedinthestar-modelwithfull-bodybeingthecenter.Sincethepartdetectionscoreisconsideredasinputofourframework,theframeworkkeepsunchangedifotherarticu-lationmodelsorfeaturesareused.FortheexperimentonthedatasetsCaltech-Train,ETHandPETS2009,theIN-RIAtrainingdatasetin[3]isusedfortrainingourpartsmodelanddeepmodels.FortheexperimentontheCaltech-Testdataset,Caltech-Traindatasetisusedfortrainingpartsmodelanddeepmodels.Intheexperiments,wemainlycomparewiththeapproachD-Isol[25].Itusesthesamefeature,thesamedeformablemodelandthesametrainingdatasetasoursfortrainingthepartsmodel.D-Isol[25]onlyusedthedeepmodelforisolatedpedestrianswhilebothiso-latedpedestriansandco-existingpedestrianareconsideredinthispaper.TheFPDWin[8]andtheCrossTalkin[7]arealsoincludedforcomparison.FPDWandCrossTalkaretrainedonINRIAtrainingdatasetusingboostingclassi eronmultiplefeatures.
Theper-imageevaluationmethodologyassuggestedin[9]isused.WeusethelabelsandevaluationcodeprovidedbyDoll´arin[9].Asin[9],log-averagemissrateisusedastheevaluationcriterion.
5.1.ExperimentalResultsonfourpubliclyavailable
datasets
Inthissection,pedestriansatleast50pixelstall,fullyvisibleorpartialoccludedareinvestigatedintheexperi-ments.Thissetofpedestriansisdenotedasthesubsetrea-sonablein[9].
2http://www.cvg.rdg.ac.uk/PETS2009/a.html
Figure6.ExperimentalresultsontheCaltech-Traindataset(left)andtheETHdataset(right)forHOG[3],LatSVM-V2[14],FPDW[8],D-Isol[25]andourmutualvisibilityapproach,i.e.D-Mut.
Fig.6showstheexperimentalresultsontheCaltech-TraindatasetandtheETHdataset.Fig.8showsde-tectionresultcomparisonofD-IsolandD-Mutat1FPPIontheCaltech-TraindatasetandtheETHdataset.Itcanbeseenthatourmutualvisibilityapproach,i.e.D-Mut,has4%and6%http://www.77cn.com.cnparedwithLatSVM-V2,ourapproachachieves11%and10%http://www.77cn.com.cnparedwithFPDW,ourapproachachieves5%and19%http://www.77cn.com.cnparedwiththeimage-basedapproachesevaluatedin[9],ourapproachhasthelowestmissrateontheETHdatasetandtheCaltech-Traindataset.
http://www.77cn.com.cnparedwithD-Isol,D-Muthas5%missrateimprovementontheCaltech-Testdatasetand8%missrateimprovementonthePETS2009dataset.Ourapproachhasthesamemissrateasthebestperformingapproaches[27,6]ontheCaltech-Testdataset,bothofwhichhave48%averagemissrate.Morediscrimi-nativefeatures[31](pyramidHOGandcolor-self-similarityfeatures)andscenegeometricconstraints[27]havebeenusedintheseapproaches.Thesefeaturesandconstraintscanalsobeusedforfurtherimprovingourresults.Forex-ample,withgeometricconstraintsusedinanunsupervisedway,themissrateofD-Mutcanbereducedfrom48%to44%.ThePETS2009crowddatasetisawell-knownbench-markforpedestriancountingandpedestriantracking.Inourexperiment,weselectS2L2withmediumdensitycrowdandS2L3withhighdensitycrowdfortest.S2L2con-tains436framesandS2L3contains240frames.Therearetotally676framesand14385pedestriansevaluatedinthisexperiment.Wemanuallylabeledthepedestriansinthisdataset3.TheresultsforLatSVM-V2andFPDWareobtainedbyrunningtheircodeforthisdataset.Theexperi-mentalresultsforCrossTalkisnotavailableonPETS2009becausethecodeisnotavailable.
3http://www.ee.cuhk.edu.hk/
xgwang/2DBNped.html