外文翻译文档格式.docx
《外文翻译文档格式.docx》由会员分享,可在线阅读,更多相关《外文翻译文档格式.docx(11页珍藏版)》请在冰豆网上搜索。
Inthispaperwepresentanewalgorithmforreal~timedetectionandtrackingofmovingtargetsinterrestrialscenesusingamobilecamera.Ouralgorithmconsistsoftwomodes:
detectionandtracking.Inthedetectionmode,backgroundmotionisestimatedandcompensatedusinganaffinetransformation.Theresultantmotionrectifiedimageisusedfordetectionofthetargetlocationusingsplitandmergealgorithm.Wealsocheckedotherfeaturesforprecisedetectionofthetargetlocation.Whenthetargetisidentified,algorithmswitchestothetrackingmode.ModifiedMoravecoperatorisappliedtothetargettoidentifyfeaturepoints.Thefeaturepointsarematchedwithpointsintheregionofinterestinthecurrentframe.Thecorrespondingpointsarefurtherrefinedusingdisparityvectors.Thetrackingsystemiscapableoftargetshaperecoveryandthereforeitcansuccessfullytracktargetswithvaryingdistancefromcameraorwhilethecameraiszooming.Localandregionalcomputationshavemadethealgorithmsuitableforreal-timeapplications.Therefinedpointsdefinethenewpositionofthetargetinthecurrentframe.Experimentalresultshave
shownthatthealgorithmisreliableandcansuccessfullydetectandtracktargetsinmostcases.
Keywords:
realtimemovingtargettrackinganddetection,featurematching,affinetransformation,vehicletracking,mobilecameraimage.
1Introduction
Visualdetectionandtrackingisoneofthemostchallengingissuesincomputervision.Applicationofthevisualdetectionandtrackingarenumerousandtheyspanawiderangeofapplicationsincludingsurveillancesystem,vehicletrackingandaerospaceapplication,tonameafew.Detectionandtrackingofabstracttargets(e.g.vehiclesingeneral)isaverycomplexproblemanddemandssophisticatedsolutionsusingconventionalpatternrecognitionandmotionestimationmethods.Motion-basedsegmentationisoneofthepowerfultoolsfordetectionandtrackingofmovingtargets.Itissimpletodetectmovingobjectsinimagesequencesobtainedbystationarycamera[1],[2],theconventionaldifference-basedmethodsfailtodetectmovingtargetswhenthecameraisalsomoving.Inthecaseofmobilecameraalloftheobjectsintheimagesequencehaveanapparentmotion,whichisrelatedtothecameramotion.Anumberofmethodshavebeenproposedfordetectionofthemovingtargetsinmobilecameraincludingdirectcameramotionparametersestimation[3],opticalflow[4],[5],andgeometrictransformation[6],[7].Directmeasurementofcameramotionparametersisthebestmethodforcancellationoftheapparentbackgroundmotionbutinsomeapplicationitisnotpossibletomeasuretheseparametersdirectly.Geometrictransformationmethodshavelowcomputationcostandaresuitableforrealtimepurpose.Inthesemethods,auniformbackgroundmotionisassumed.Anaffinemotionmodelcouldbeusedtomodelthismotion.Whentheapparentmotionofthebackgroundisestimated,itcanbeexploitedtolocatemovingobjects.Inthispaperweproposeanewmethodfordetectionandtrackingofmovingtargetsusingamobilemonocularcamera.Ouralgorithmhastwomodes:
detectionandtracking.Thispaperisorganizedasfollows.InSection2,thedetectionprocedureisdiscussed.Section3describesthetrackingmethod.ExperimentalresultsareshowninSection4andconclusionappearsinSection5.
2Targetdetection
InthedetectionmodeweusedaffinetransformationandLMedS(Leastmediansquared)methodforrobustestimationoftheapparentbackgroundmotion.Afterthecompensationofthebackgroundmotion,weapplysplitandmergealgorithmtothedifferenceofcurrentframeandthetransformedpreviousframetoobtainanestimationofthetargetpositions.Ifnotargetisfound,thenitmeanseitherthereisnomovingtargetinthesceneor,therelativemotionofthetargetistoosmalltobedetected.Inthelattercase,itispossibletodetectthetargetbyadjustingtheframerateofthecamera.Thealgorithmaccomplishesthisautomaticallybyanalyzingtheproceedingframesuntilamajordifferenceisdetected.Wedesignedavotingmethodtoverifythetargetsbasedonaprioriknowledgeofthetargets.Forthecaseofvehicledetectionweusedverticalandhorizontalgradientstolocateinterestingfeaturesaswellasconstraintonareaofthetargetasdiscussedinthissection.
2.1Backgroundmotionestimation
Affinetransformation[8]hasbeenusedtomodelmotionofthecamera.Thismodelincludesrotation,scalingandtranslation.2~Daffinetransformationisdescribedasfollow:
(1)
where(xi,yi)arelocationsofpointsinthepreviousframeand(Xi,Yi)arelocationsofpointsinthecurrentframeanda1~a6aremotionparameters.Thistransformationhassixparameters;
therefore,threematchingpairsarerequiredtofullyrecoverthemotion.Itisnecessarytoselectthethreepointsfromthestationaryback~groundtoassureanaccuratemodelforcameramotion.WeusedMoravecoperator[9]tofinddistinguishedfeaturepointstoensureprecisematch.Moravecoperatorselectspixelswiththemaximumdirectionalgradientinthemin~maxsense.
Ifthemovingtargetsconstituteasmallarea(i.e.lessthan50%)oftheimage,thenLMedSalgorithmcanbeappliedtodeterminetheaffinetransformationparametersoftheapparentbackgroundmotionbetweentwoconsecutiveframesaccordingtothefollowingprocedure.
1.SelectNrandomfeaturepointfrompreviousframe,andusethestandardnormalizedcrosscorrelationmethodtolocatethecorrespondingpointsinthecurrentframe.Normalizedcorrelationequationisgivenby:
(2)
here
and
aretheaverageintensitiesofthepixelsinthetworegionsbeingcompared,andthesummationsarecarriedoutoverallpixelswithinsmallwindowscenteredonthefeaturepoints.Thevaluerintheaboveequationmeasuresthesimilaritybetweentworegionsandisbetween1and-1.Sinceitisassumedthatmovingobjectsarelessthan50%ofthewholeimage,thereforemostoftheNpointswillbelongtothestationarybackground.
2.SelectMrandomsetsofthreefeaturepoints:
(xi,yi,Xi,Yi)fori=1,2,3,fromtheNfeaturepointsobtainedinstep1.(xi,yi)arecoordinatesofthefeaturepointsinthepreviousframe,and(Xi,Yi)aretheircorrespondsincurrentframe.
3.Foreachsetcalculatetheaffinetransformationparameters.
4.TransformNfeaturepointsinstep1usingMaffinetransformations,obtainedinstep3andcalculatetheMmediansofsquareddifferencesbetweencorrespondingpointsandtransformedpoints.Thenselecttheaffineparametersforwhichthemedianofsquareddifferenceistheminimum.
Accordingtotheaboveprocedure,theprobabilitypthatatleastonedatasetinthebackgroundandtheircorrectcorrespondingpointsareobtainedisderivedfromthefollowingequation[7]:
(3)
where
(<
0.5)istheratioofthemovingobjectregionstowholeimageandqistheprobabilitythatcorrespondingpointsarecorrectlyfind.In[7]ithasbeenshownthattheabovemethodwillgiveanaccurateandreliablemodel.
2.2Movingtargetdetectionusingbackgroundmotioncompensatedframes
Whenaffineparametersareestimated,theycanbeusedforcancellationoftheapparentbackgroundmotion,bytransformationofpreviousframe.Nowdifferenceofthecurrentframeandtransformedpreviousframerevealstruemovingtargets.Thenweapplyathresholdtoproduceabinaryimage.Theresultsofthetransformationandsegmentationareshownisfigure1~aand1~b.Somepartsaresegmentedasmovingtargetsduetonoise.Connectedcomponentpropertycanbeappliedtoreduceerrorsduetonoise.Weusesplitandmergealgorithmtofindtargetbounding-boxes.Ifnotargetisfound,thenitmeanseitherthereisnomovingtargetinthesceneor,therelativemotionofthetargetistoosmalltobedetected.Inthelattercase,itispossibletodetectthetargetbyadjustingtheframerateofthecamera.Thealgorithmaccomplishesthisautomaticallybyanalyzingtheproceedingframesuntilatargetisdetected.Ourspecialinterestisdetectionandtrackingofthemovingvehiclessoweusedaspectratioandhorizontalandverticallineasconstraintstoverifyvehicles.Ourexperimentsshowthatcomparisonofthelengthofhorizontalandverticallinesinthetargetareawiththeperimeterofthetargetwillgiveagoodclueaboutthenatureofthetarget.
3Targettracking
Afteratargetisverified,thealgorithmswitchesintothetrackingmode.ModifiedMoravecoperatorisappliedtothetargettoidentifyfeaturepoints.Thesefeaturepointsarematchedwithpointsintheregionofinterestinthecurrentframe.Disparityvectorsarecomputedforthematchedpairsofpoints.Weuseddisparityvectorstorefinethematchedpoints.Therefinedpointsdefinethenewpositionofthetargetinthecurrentframe.Thealgorithmswitchestothedetectionmodewheneverthetargetismissed.Althoughthedetectionalgorithmdescribedabovecanbeusedfortrackingtoobutthetrackingalgorithm,wedescribeinthissectionhasverylowcomputationcostincontrastwiththedetectionalgorithmdescribedabove.Ontheotherhandwhenthetargetisdetecteditisnotrestrictedtokeepmovingintrackingmode.Thetargetcanalsobelargerthan50%ofthesceneryinthetrackingmodeandthismeanscameracanzoomtohavealargerviewofthetargetwhiletracking.
Figure1:
twoconsecutiveframesanddifferenceofthemafterbackgroundmotioncompensation,thecalculatedaffineparametersare:
a1=0.9973,a2=-0.004,a3=0.008,a4=1.0022,a5=1.23,a6=-2.51
Whenthesizeofthetargetisfixedtheno