22-03-2012, 12:19 PM
A shape-based voting algorithm for pedestrian detection and tracking
A shape-basedvotingalgorithmforpedestriandetectionandtracking.pdf (Size: 1.16 MB / Downloads: 49)
Introduction
Pedestrian trackingisachallengingproblem,withanumberof
application areasincludingautomatedsurveillance.Mostcurrent
techniquesusebackgrounddifferencing,whichiscomputationally
efficient butfailsinscenesthataretoomobile,orwherethecamerais
non-stationary.Thisincludessceneswithdense,wind-blownfoliage,
water, andbackgroundtraffic.Shape-based trackingalgorithms
provideapromisingalternative,beinginsensitivetotheseproblems.
This paperpresentsanewshapemodel,theMOUGH(mixtureof
uniform andGaussiansHough)transform,basedonaGaussian
mixture model(GMM)representationoforientededgellocations.It
is trainedfromsampleimages,andthenmaybeusedtodetectand
track theobjectclass.Wecomparethenewalgorithmagainst
earlier variantsofthegeneralizedHoughtransform,andactive
shape models.Sincethemethodisshape-based,itdoesnotrelyon
movement, andcanalsobeusedtodetectstationaryobjects.
This paperisstructuredasfollowsectiontwodiscussesshape-
based detectionandtracking,sectionthreeintroducestheMOUGH
transform, sectionfourdiscussesMOUGHfortracking,sectionfive
presents results,andsectionsixconcludesthepaper.
TheMOUGHtransform
. AGMM-basedvotingalgorithm
This sectionintroducestheMOUGH(mixtureofuniformand
Gaussian Hough)transform;see Fig. 2. LikeEcabertandThiran’s
approach [26], theR-Tablebecomesamappingfromedgeorienta-
tion toaprobabilitydistributionovercentroidoffsets.However,
whereas EcabertandThiranuseanon-parametricapproach,weuse
a GaussianMixtureModel(GMM).Thesemi-parametricrepresen-
tation ismorecompactthanthenon-parametricone,allowingusto
interpolate betweenarelativelysmallnumberofexemplarimages,
ignoring outlyingpointsandbridginggaps.
Detectingandtrackinganindividual
toprow,showsthemarginaldistribution, pðdjFGÞ, forfour
of theMOUGHmodelsontrainingsetone.Eachmodelhasfive
Gaussian componentspermixture.Theeffectivenessofcentroid
adjustment intighteningthemodel,oftheoutlierdistributionin
shedding non-essentialedges,andthebenefitofcombiningthe
two, isclear.Sinceeveryimagerepresentsthesameindividual,
quite alargeoutlierdistributionwasused(wout ¼ 0:2), inferringa
very tightdistribution.Incontrast,averysmalloutlierdistribution
(wout ¼ 0:01) isusedontrainingsettwo,toenhancegeneralization.
We alsousealargernumberofcomponents(20)permixture,
as themoregeneraldatasetcontainsmoresubtleties(see Fig. 7,
bottom row);andscale-adjustment,sincetheprototypesvaryin
scale—this significantlyenhancesperformance.