DOC

Detecting

By Lewis Rodriguez,2014-07-06 14:46
7 views 0
Detecting

    Detecting

    Trans.TianjinUniv.2009.15:013-018

    DOl10.1007/s12209.009.00038

    DetectingandTrackingMovingTargets

    on0mnidirectionalVision

    YANGShuying(杨淑莹),GEWeimin(葛为民)2ZHANGCheng(张成)

    HEPilian(何丕廉)

    (1.SchoolofComputerScienceandTechnology,TianjinUniversity,Tianjin300072,China; 2.TianjinKeyLabofIntelligentComputingandNovelSoftwareTechnology,

    TianjinUniversityofTechnology,Tianjin300191,China)

    Abstract:Amethodwaspresentedtoimplementthedetectingandtrackingofmovingtargetsthroughomnidirec.

    tionalvision.Themethodcombinedopticalflowwithparticlefilterarithmetic,inwhichopticalflowfieldwasused

    todetectandlocatemovingtargetsandparticlefilterwasusedtotrackthedetectedmovingobjects.Accordingto

    thecircularimagecharacterofomnidirectionalvision,thecalculationequationofopticalflowfieldandthetracking

    arithmeticofparticlefilterwereimprovedbasedonthepolarcoordinatesattheomnidirectionalcenter.

    Theedgeof

    arandomlymovingobjectcouldbedetectedbyopticalflowfieldandwassurroundedbyareferenceregioninthe

    particlefilter.Adynamicmotionmodelwasestablishedtopredictparticlestate.Histogramswereusedasthefea

    turesinthereferenceregionandcandidateregions.Themutualinformation(MI)andGaussianfunctionwerecorn

binedtocalculateparticleweights.Finally,thestateoftrackedobiectwascomputedbythetota

    lparticlestateswith

    weights.Experimentresultsshowthattheproposedmethodcoulddetectandtrackmovingobj

    ectswithbetterrea1..

    timeperformanceandaccuracy.

    Keywords:omnidirectionalvision;opticalflow;particlefilter;mumalinformation

    Detectingandtrackingobjectsisachallengingprob

    lemduetothepresenceofnoise,clutteranddynamic changesinascene.Theaimofdetectingandtrackingis toextractmovingtargetsinformationfromcomplex background,selectpropertargetfeatures,andtrack1110V

    ingtargetsrealtime.Thearithmeticofdetectingand trackinghasbeenproposedandimplementedtoover

    comethesedifficulties.

    Opticalflowiscausedbytherelativemotionbe

    tweentargetsandacamera.Opticalflowfieldpresents thedistributionofapparentmovingvelocitiesinabright

    nessimage.Theinformationofdenselymovingpoints canbeobtainedbyopticalflowfield.Opticalflowex

    pressesthevariationofbrightnessinformationofmoving targets,anditcanbeusedtodetectanobjectmo

    tion[".Thisapproachneednotdefine~aturesofamoving targetinadvance.Opticalflowhasbeenextensivelystud- iedinthecomputervision.Haetalproposedtodetect flyingvehicle'smotionbyopticalflowarithmetic.Ya

    mamototal[3]andTsutsuiPtl[4]detectedmultiplemov.

    ingobjectsbasedonopticalflowbyusingmultiplecam- erasinindoorenvironments.However.opticalflowcom. putationneedsmuchtime.itjsnotsuitableforrealtime

    tracking.

    Particlefilter(PF1hasalsobeenwidelyusedin visualtrackingfield.Itoffersaprobabilisticframework fordynamicstateestimation.ItissuccessfuIfornonlin

    earnon.Gaussianestimationproblems.Systemstatetran

    sitionsexpressthemodelofmovingtargetsinparticle filter.Theflexibilityisimprovedinparticlefiltercorn. paredwiththeKalmanfilter.Someparticlefiltersbased onobjecttrackingmethodshavebeenproposedinthe literature.Perezeti[5]andNummiaro,[6]prOpOseda

    colorprobabilistictrackingmethod,whichcomparedthe colorcontentofcandidateregionswithareferencere. gion,andreachedapproximationoftheposteriorbypar

    ticlefilter.GuoandQinproposedtouseparticlefilter fortrackinggroundmovingtargets.

    Inoursystemanomnidirectionalcamerawithafish eyewasused.Theomnidirectionalvision(omnivision)

    canprovidea360.viewoftheenvironmentinasingle imageandhasbeenappliedinthefieldofcomputervi

    sioninrecentyears.Formeomnidirectionadvantagesof compactvisualinformationanddirectionfeatures.ithas Accepteddate:20071115.

    SupportedbyTianjinHigherEducationTechnologyDevelopmentFoundation(No.200713

    08),TianjinNaturalScienceFoundation(06YFJMJC03600)and NationalNaturalScienceFoundationofChinarNo.60773073) YANGShuying,bornin1964,female,Dr,Prof.

    c0"esp0ndencetoYANGShuying,Email:ysyingl26@126corn

    TransactionsofTianjinUniversityVo1.15No.J2009 greatpromisefortargettrackingandsuitswiderange applications.suchasautonomousnavigationscene reconstructiOnandsurvelllance[91.

    Inthispaperanintegratedmethodusingopticalflow andparticlefilterwaspresentedtoimplementdetecting andtrackingofmovingtargets.Omnidirectionalimage fomniimage)hasadditionaldistortionsonthereflected surface.Asaresult,thetraditionaltrackingarithmetic couldnothaveagoodperformanceinaccuracyand speed.Accordingtothecircularimagecharacterofthe omniimage,polarcoordinateswereconstructedtode

    creasethenumberofpixels.Thecalculationequationof opticalflowfieldandthetrackingarithmeticofparticle filterwereimprovedbasedonthepolarcoordinates. Thus.theproposedmethodsuitsomnidirectionalvision system.

    1Detectingmovingtargetsusingopticalflow arithmetic

    ByusingaCCDcamerasystem,themovingobjects areprojectedona2Dimageplane.2Dvectorfieldnamed motionfieldcanbeacquired.Themotionestimationis calledODticalflow[

    ,

    whichisadirectionvectorfield.

    Therelativemotionvectorcanbemeasuredfromthe instantaneouschangesofbrightnessvalueatapixel point.Inordertocalculateopticalflowfield,itmaybe assumedthattheobjectbrightnessischangelesswhenan objectmoves.GivenabrightnessfunctionI,Y,t)ata pixelposition,Y)andtimet,thisbrightnessconserva. tionconditioncanbewrittenas

    I(x,Y,t)=I(x+Vxt,Y+vvt,0)(1)

    where1,andv,,aredefinedasthemotionvelocitycom-

    ponentsinxandYdirection,respectively.Differenceof I(x,Y,t)withrespecttotisassumedaszero,givenas :++

    :0f21

    dtdxdtdydtdt

    1.1Constructingpolarcoordinatesinanomni- image

    AnomniimageisshowninFig.1.Itlookslikea

    circle.Eachpixelintheomni-imagehasitsbrightness. Itsbrightnesscanbeexactlyacquiredbasedoneither XYcoordinatesorpolarcoordinates.Sincetheoptical flowcomputationcoststoomuchtime,afastmethodis presented.

    Thecentero(0,0)inxYcoordinatesislocatedon thelefttop.Butthecenter0,R)inpolarcoordinatesis locatedatthecenterofanomni-image(seeFig.2).The center0(,R)intheomniimagecanbeobtainedby

    calculation.Firstly,XYcoordinatesaretransformedinto Ycoordinates.P,Y)isthepixelbasedonthecenter o(0,0).P,y)canbetransformedintoP(x,Y), whichisbasedonthecenter0(,R)ofomni-image. Thepolarcoordinatescanbecalculatedbythefollowing equations:

    r=x+Y(3)

    0=arctan(y/x)(4)

    where0f0.?0<360.1istheanglebetweenOand

    Xaxis,andisthelengthofO'P(0?r?R).Sothe

    pixelP,Y)canbeexpressedasP(x,Y)ore(r,). Fig.1Omnidirectionalimage

    y

    Fig.2Sketchmapofcoordinatetransformation 1.2Calculatingthevelocityfieldanddetecting movingtargets

    Opticalflowofeverypixelintheomni-imageneed notbecalculated.Somepixelshavebrightnessinpolar coordinates.Thelineofevery1.canbepaintedfromthe centertotheboundaryinomniimage.Thelengthofthe

    1ineisfrom0toR.Theangle0ofthelineisfrom0.to 359..Sothetotalnumberofpixelsis360xR.Theopti

    calflowsofthesepixelsarecalculated.Thus.thenumber ofpixelsisdecreasedgreatlycomparedwiththeoriginal image,thetimeofcomputingopticalflowfieldisshort, andthespeedofrecognizingmovingtargetsisimproved. Inthispaperthe3DSobelconvolutionkemelis YANGShuyingetalDetectingandTrackingMovingTargetsonOmnidirectionalVision

    ThesumofIr,

    Io,Itcanbecalculatedby3DSobel

    31utionkernelfromthelastflame,currentframe, extflame.

    rheparameter"isthecentripetalvelocityfrom :center;1,

    ,

    istheanglevelocity.Thefollowing

    flowfieldcanbelocated.Thus,theopticalflowtech

    niquecanbeusedtoobtaintheinitialtargetregion.The regionisdefinedasthereferenceregionintheparticle filter.Thenextstepistousetheparticlefiltertotrackthe movingobjects.

    ionsareusedtocalculatethevelocityfield

    ]foreachpixelP(r,)inthecuHentameat2Trackingmovingtargetsusingparticle

c0ordinates:filter

    VTM

    (5)2.1Particlefilter

    nEqs.(5,6),[uk,1,kisthevelocityestimatefor

    ,ixelP(r,0),and["--k,]istheneighborhood

    geof[",V.Theinitialkis0.Theparameters

    .

    areinitializedas0foreachpixelinthecurrent :atpolarcoordinates.Theaveragecentripetalveloc

    foreachpixeliscomputedalongthedirectionof l,,r1)byusing[111;101;111]askerne1. tverageanglevelocityk

    ,

    foreachpixelalongthe

    :ionof(+1,0,01)byusing[101;101;

    ]askerne1.Theparametero~isthesmoothnessfac- ere=500.

    FhepicturesinFig.3showthatopticalflowfield ecomputedbyusingtheseframes.Themovingob opticalflowfieldisdifferentfromtheothersin ~roundfromFig.3(d).Themovingtargetinoptica1 (b)Thecurrentframe

    (c)Thenextflame(d)Opticalflowfield Fig.3Sketchmapofcomputingopticalflow Particlefilterisasampleprocessingmethodwhich isbasedonthechangeofBayesfilter.Theweightsofa finitesamplesset()aredelegatedtotheapproximate posteriordistribution.

    :

{(,?I)li=1,,n}(7)

    wheresexpressesthestateofaparticle;isanon

    negativenumericalfactorcalledimportanceweight,and thetotalofisone.

    Theparticlefilterisdividedintotwosteps,predic

    tionandupdate.

    Uptotimet1,allavailableobservations,z11= {z1,,z1},aregiven.Thepredictionstatep(s{zl:_1)

    atthenexttimetbyusingtheprobabilistictransition modelis

    p(slzn1)=Ip(slst1)p(f-lZl:t-I)dsf-(8)

    Attimet,theobservationzfisavailableandthe statecanbeupdatedbyusingBayes'rule. ,)=

    p(zzIt-1

    (9)

    ,l

    Here,theposteriorp(z,lsf)isdescribedbytheobserva

    tionequation.

    Intheparticlefilter,p(slz?)isapproximatedbya

    finitesetofnsamples{:.

    withimportanceweights

    {},=l.Thecandidatesamplesaredrawnfroman importancedistributionq(sISl:t-Izh).Theweightsofthe samplesarecomputedas

    Inordertoavoiddegeneracy,thesamplesneedto beupdatedaccordingtotheirimportanceweights.Inthe caseofboostingfilter,q(s,ls_I'zn)=p(sI,and theweightsbecometheobservationlikelihoodp(zl), =

    ,p(zls).

    2.2MIfeatures

    MIisusedtoestimatethesimilaritybetweenrefer

    enceregionandcandidateregion.Histogramisusedto lJ

    TransactionsofTianjinUniversityVo1.15No.j2009 computeMI.Theimplementingstepsareasfollows. (1)Obtainhistogramarraysofreferenceregion [],candidateregionHb[],andjointhistogram arrayHa^[,K],Kisthebinsofhistogram. (2)Calculatetheprobabilityofreferenceregion, candidateregion,andjointarray.

    []=(11)

    p:

    ??h]i=0j=0

    i=0,,K;j=0,,K

    (3)Calculateentropyofreference

    region,andjointarray.

    =

    ?([i1?log(p?)i=o

    =

    ?(log(p)O

    Al,A2,A3,A4areconstants;B1,B2,B3,B4representparticle propagationradius;

    representsarandomnumberin

    [1,1].

    Toassignparticleweight,Giscomputedbetween eachparticleandtargettemplateusingEq.(17).Sothe particlesfromtargettemplatewithsmallGIcanbe

foundandaregivenlessweightscomparedwiththepar

    ticleswithlarger1.Theweightsareassignedusing p(z,I=-~-~exp{一击_}c20,)={一击ll}()

    f13)contant.

    ?I=p(zli=1,,n(21)

    Iftheweightofaparticleistoosmal1.theparticle isresetusingtheresamplingtechnique.Here,thepa

    region,candidaterameterswithsmallerweightparticlesarecoveredbythe

    parameterswiththelargestweightparticle. f141Afternormalizingtheparticleweights,theparame- tersoftrackedtargetarecomputedandtheyarethetotal f151particlestateswithweights.

    ?曲=一??(f,log(p(16)

    i=oj=o

    (4)CalculateGMIbetweenreferenceregionandcan

    didateregion.

    GMI=H+HbH(17)

    2.3Trackingmovingtargetsusingparticlefilter Heretheinitialreferenceregionisdefinedasarec

    tangletemplateandtheinitialmotionparameters (r,0,,,h)ofreferenceregioncanbeacquired.Thepa- rameters(,0)representthepositionatpolarcoordi- natesbasedonthecenterO(,R)ofanomni-imageand (,,h)representthewidthandheightoftherectangle template.

    ParticlesetS={((,,,,h),14,)}1,,}isran

    1

    domlydistributedinwhich{w=Ii=1,,n}andn

    n

    representsthenumberofparticles.

Report this document

For any questions or suggestions please email
cust-service@docsford.com