发表在CVPR13上的行人检测方法
Table1.Overviewofourpedestriandetectionapproach.1.obtainpartdetectionscoresbypartdetector;
2.estimatep(y1|y2=0,x1)in(2)andφ(y;x)in(3)withthedeepmodelinSection4,estimateφ|xp(y;x)in(3)withGMM;3.p(y11,x2)=p(y1,y2=0|x1,x2)+p(y1,y2=1|x1,x2).
modelwasusedforpedestriandetectionin[25,24].The
approachesin[25,24,19]focusedonisolatedobjectsorpedestrians.Thispaperfocusesonco-existingpedestrians,whichhasnotbeenconsideredin[25,19,24].
3.Overviewofourapproach
Inthispaper,wemainlydiscusstheapproachforpair-wisepedestriansandextendittomorepedestriansinSec-tion4.3.Denotethefeaturesofdetectionwindowwndbyvectorx1
1,containingbothappearanceandpositionin-formation.Denotethelabelofwnd1byy1∈{0,1}.Pedestriandetectionwithadiscriminativemodelaimsatobtainingp(y1|xfor1)foreachwindowwndallsizesofwindows.1inaslidingwin-dowmannerWeconsideran-otherdetectionwindowwnd∈{0,1}.Andwehavethe2withfeaturesxfollowingbymarginalizing2andlabelyy22:
p(y
1|x1,x2)=
p(y1,y2|x1,x2)y2=0,1(1)
=p(y1,y2=1|x1,x2)+p(y1,y2=0|x1,x2),Wheny2=0,wehave
p(y1,y2=0|x1,x2)=p(y1|y2=0,x1)p(y2=0),(2)wherep(y1|y2=0,x1)isobtained0)isfromaconstantthedeeppriormodelforisolatedpedestrians.p(ywhich2=onwnd2beingabackground,isobtainedfromcross-validation.Wheny2=1,wehave
p(y1,y2=1|x1,x2)∝φ(y;x)φp(y;x),
(3)
φ(y;x)in(3)isusedforrecognizingpair-wiseco-existing
pedestriansfrompartdetectionscores,wherex=[xT1xT2]T
,y=1ify1=1andy2=1,otherwisey=0.Bothp(ydeep1|ymodel2=0,xintroduced1,x2)andφ(y;x)areobtainedfromtheinSection4.φtherelativepositionp(y;x)in(3)mod-elsprobabilityforbetweenwndwndφmixturemodel1and(GMM).2.p(y;x)isestimatedfromGaussianAnoverviewofourapproachisgiveninTable1.
4.Themutualvisibilitydeepmodel
Sincethevisibilityrelationshipofpartsbetweenpair-wisepedestriansisdifferentwhenpedestrianshavedifferentrelativepositions,therelativepositionsareclusteredintoKmixturesusingGMM.AndKdeepmodelsaretrainedfortheseKmixtures.Apairofdetectionwindowsare
(a)
(b)
Figure2.(a)Themutualvisibilitydeepmodelusedforinferenceand netuningparametersand(b)thedetailedconnectionandpartsmodelforpedestrian1.
classi edintothekthmixtureandthenthispairareusedbythekthdeepmodelforlearningandinference.Thedifferencesbetweenthetwopedestriansinhorizontallo-cation,verticallocationandsize,denotedby(drandomvariablesintheGMMx,ddistribu-y,ds),areusedasthetionp(dx,dy,d.φs).Positivesamplesisobtainedarefromusedp(fordtrainingp(dx,dy,ds)p(y;x)in(3)x,dy,ds).
4.1.Thedeepmodelattheinferencestage
Fig.2(a)showsthedeepmodelusedattheinferencestage.Fig.2(b)showsthepartsmodelusedforpedestrian1atwindowwnddowwnd1.Thepartsmodelforpedestrian2atwin-2isthesame.AsshowninFig.2(b),thereare3layersofpartswithdifferentsizes.Foreachpedestrian,therearesixsmallpartsatlayer1,sevenmedium-sizedpartsatlayer2andsevenlargepartsatLayer3.Thesixpartsatlayer1areleft-head-shoulder,right-head-shoulder,left-torso,right-torso,left-legandright-leg.Apartatanup-perlayerconsistsofitschildrenatthelowerlayer.Thepartsatthetoplayerarethepossibleocclusionstatuseswithgraycolorindicatingocclusions.
ThedetectionscoresforLlayersaredenotedbys=
[s1T
...sLT]T=γ(x),whereγ(x)isobtainedfrompartdetectors,slforl=1,...,Ldenotesthescoresatlayerl.ForthemodelinFig.2,L=3.Andwehave
sl=[slT1slT2]T,wherethePl
scoresofthetwopedestrians
atlayerlaredenotedbysl1=[s11,1,...,slT
1,Pl]andsl2=
[s12,1,...,slT
2,Pl].ThevisibilitiesofPlpartsaredenotedby hl=[h1,...,hll]Tandspectively.11,1h l2=[h12,...,hll]Tre- hl=[h lT1 ThehThidden1,Pvariablesatlayer,1laredenoted2,P
by
l2
]T.Sinceh lisnotprovidedattrainingstage