发表在CVPR13上的行人检测方法
ortestingstage,itisconsideredasahiddenrandomvector.Inourimplementation,DPMin[14]isusedforobtainingpartdetectionscoresins.Thedeformationamongpartsarearrangedinthestar-modelwithfull-bodybeingthecenter.Inthispaper,itisassumedthatpart-basedmodelshavein-tegratedbothappearanceanddeformationscoresintos.Inordertohavethetoplayerrepresentingocclusionstatusinamoredirectway,s3accumulatethedetectionscoresthat ttheirpossibleocclusionstatuses.Forexample,
s31,1= s31,1+s21,1+s11,1+s1
1,2,
s31,2
=
s 3 41,2
+
s2 41,i
+
s1(4)
1,i,
i=1
i=1
wheresl1,iforl=1,2,i=1,...,Pl
isthedetectionscorefortheithpartatlayerl,s 31,1isthedetectionscoreforthehead-shoulderpartatlayer3ands 3forthehead-torsopartatlayer3.1,2isthedetectionscoreInourimplementationofthedetector,thehead-shoulderpartatthetoplayerhashalfoftheresolutionofHOGfeaturescomparedwiththehead-shoulderpartatthemiddlelayer.
Theoverlapinformationatlayer2inFig.2isdenotedby
o=[oTTheoverlap1oT2]T,whereoinformationn=[oforsixn,1opartsn,2...oareleft-head-shouldern,6]T
forn=1,2.on,1,right-head-shoulderon,2,left-torsooleft-legoInordern,3,right-torsoon,4,n,5andright-legon,6.toobtaino,theoverlapofthesesixpartswiththepedestrianregionoftheotherpedestrianiscomputed.Accordingtotheaver-agesilhouetteinFig.3(a),whichisobtainedbyaveragingthegradientofpositivesamples,tworectanglesareusedforapproximatingthepedestrianregionoftheotherpedestrian.Onerectangleisusedfortheheadregion,denotedbyAisusedforthetorso-legregion,denotedh,anotherrectanglebyAt.Denotetheregionon,ibyAn,i.on,iisobtainedasfollows:
on,i=
area(An,i∩Ah)+area(An,i∩At)
area(A,
(5)
n,i)
wherearea(·)computestheareainthisregion,∩denotesintersectionofregion.Forexample,therightpersoninFig.3(b)hastheleft-head-shoulder,left-torsoandleft-legoverlappingwiththepedestrianregionsoftheleftperson.SinceAn,i,AhandAarea(·)and∩intarerectangularregions,theopera-tions(5)canbeef http://www.77cn.com.cnparedwithsegmentation,therectangularregionisanapproximatebutfasterapproachforobtainingpedestrianre-gionandcomputingtheoverlapinformationo.
Attheinferencestage,thepedestrianco-existencelabelyisinferredfromfeaturesx.Thepartvisibility
probability
Figure3.(a)Tworectangular īregionsusedfor ī
approximatingthe
pedestrianregionand(b)anexamplewithleft-head-shoulder,left-torsoandleft-legoverlappingwiththepedestrianregionsoftheleftperson.
hlj
+1isobtainedusingthemodelinFig.2,i.e. hlj+1=p(hlj
+1=1|hl,x)=σ(hlT
wl
,j+cl+1+glT
jj+1slj+1),
(6)
hl= h
lifl=L 1,hl=[h lToT]T
,ifl=L 1,whereσ(t)=(1+exp( t)) 1isthelogisticfunction.The
estimatedoutputφ(y;x)isobtainedasfollows:
φ(y;x)=ey(wLT
h
L+b)
/Z,(7)
whereZ= T L
y=0,1ey(wLh+b).ForthemodelinFig.2,
wehaveL=3.Thelearningofparameterswl
,j,wL,clj
+1andglj+1
in(6)and(7)areexplainedinSection4.2.
4.2.Thelearningofthedeepmodel
Thefollowingtwostagesareusedforlearningthepa-rametersin(6)and(7).
Stage1:Pretrainparameterswl
l+1andglFine-tunealltheparameters ,j,cStage2:j
bybackpropagat-j+1in(6).ingerrorderivatives.Thevariablesarearrangedasaback-propagation(BP)networkasshowninFig.2(a).
Asstatedin[12],unsupervisedpretrainingguidesthelearningofthedeepmodeltowardsthebasinsofattrac-tionofminimathatsupportbettergeneralizationfromthetrainingdata.Therefore,weadoptunsupervisedpretrainingofparametersatstage1.Thegraphicalmodelforunsu-pervisedpretrainingisshowninFig.4.Theprobabilitydistributionofp(h1,...,hL |x)ismodeledasfollows:
L 2p(h1,...,hL|x)=
p(hl|hl+1,x)p(hL 1,hL|x),
l=1
p(hli=1|hl+1,x)=σ(wli, hl+1+glisli+cli),
p(hL 1,hL|x)=p(hL 1 ,hL|s)=e
hL 1T
WL 1hL+(cL 1+gL 1 sL 1)ThL 1+(cL+gL sL)ThL
,(8)
where denotestheentrywiseproduct,i.e.(A B)Ai,j=i,jBi,j,hisde nedin(6).ForthemodelinFig.4,