搜索
InternalReport98{06VehicleDetectioninTra cScenesUsingShadowsbyChristosTzomakasandWernervonSeelenRuhr-Universit atBochumInstitutf urNeuroinformatik44780BochumIR-INI98{06August1998ISSN0943-2752c 1998Institutf urNeuroinformatik,Ruhr-Universit atBochum,FRGVehicleDetectioninTra cScenesUsingShadowsChristosTzomakas andWernervonSeelenInstitutf urNeuroinformatik,Ruhr-Universit atBochum,FRGAbstractInthisreportwepresentthe rstresultsofaverye cientschemeforvehicledetectionintra cscenes.Itisbasedonthedetectionoftheshadowsunderneaththevehicles.Theuseofshadowassignpatternwas rstreportedin[5].Hereinahistogramevaluationofthepavedroadsustainsthespeci cationofathresholdvaluefortheshadedareas.Thisthresholdalongwithedgeinformationleadstotheshadowdetection.Incorporatingexternalknowledgefromthe3Dgeometryandapplyingnaturalconstraintstotheclusteredsegmentsweareabletorobustlygenerateinitialobjecthypotheses.1IntroductionFortherealizationofadriverassistancesystemseveralmoduleshavetobeincorporatedatvariouslevels.Forvision-basedapproachesmaintaskslikedetectionandrecognitionofforegoingvehicleshavetobesolved.Thetimee ciencyiscrucialforthesetasks,andthusaconstrainingfactor.Afundamentalroleinthewholesystemplaystheinitialsegmentationwhichfeedslaterprocesseslikeobjectclassi cationandobjecttrackingwithinput.Asinitialsegmentationisconcernedtheextractionofsomedistinctregionsoftheimagewherethefurtherattentionhastobefocused.Ithastooperateonashort-termscaleandtosupplythesystemwithinformationaboutthecurrentsituation.Thebetterandmoreaccuratetheinitialguesses,thebetterforthegoodnessandthee ciencyoftherecognitionprocess.In[3]wehavealreadyproposedamethodforgeneratingobjecthypothesesinhighwayscenes.Themethodisbasedontextureanalysisandmoreparticularlyonimageentropy[4]andusageofthecamerageometry.Likealltexturebasedmethodsthecalculationoftheimageentropyisatime-demandingprocess.Furthermoreaseparateprocessingoffar-andnear-rangesisrequired.In[7]wehavealsoexaminedamodel-basedapproachforvehiclerecognition,whichappliesonlyfortherecognitionofacertaintypeofvehicleatatime.Di erentcategoriesofvehiclesrequiretheuseofseparatemodels.Inthisreportwesystematicallyexamineanotherfeatureusefulforvehicledetection,namelytheshadow.Indeed,theshadowunderneathvehiclesisanattributecommonforalltypesof4-wheeledvehiclesandthusitcanbeusedfortheirdetection.Weproposeamethodforextractingtheshadedsegmentsfromanimageandcombiningthemintoclusterswhichwillserveasinitialobjecthypothesesthatwillbeveri edbylater,moresophisticatedprocesses.Theapproachprovesrobustnessinvariousweatherandlightconditionsandise cientenoughforreal-timepurposes.2DetectingShadowsIn[5]itisinitiallyshownthatshadowcanbeusedasasignpatternforthedetectionofvehicles.Theauthorshaveinvestigatedtheintensityoftheareaunderneathavehicleandfoundoutthatitisdistinctlydarkerthananystainofasphaltpavedroad.In[1]we nda rstattemptthatmakesuseoftheideaofMori.Thomaneketalused30windowsscanningforhorizontaledgeelementsandappliedsomebrightness-andcorrelation-value-constraintsforthedetectionofshadowsbeneath email:Christos.Tzomakas@neuroinformatik.ruhr-uni-bochum.de12IR-INI98{06,c 1998Institutf urNeuroinformatik,Ruhr-Universit atBochum,FRGvehicles.Neverthelesstheyneithergiveanyfurtherspeci cationoftheseconstraints,nordotheydeterminehowonecanchoosethethresholdvaluefortheshadow.Theproblemisthattheintensityoftheshadowunderneathavehicledependsontheilluminationoftheimage,whichinturndependsonthedaylightandthusisnotconstant.Forthisreasonwecannotusea xedthresholdforsegmentingtheshadedareasintheimage.Furthermore,thetaskofseparatingthedarkareasinanimagefromthebrightonesisnottrivialatall.Darkorlightdarkobjectsmaybepresentintheimagewhichmakesthedetectionofavalleyinthehistogram(whichwouldseparatetheshadowsfromthebrightersurround)fromverydi cultuptoimpossible.Wedepicttheideaandtheproblemofchoosingathresholdintra csceneswiththesampleimagesin g.1.Fromlefttoright:theimageandsuccessivesegmentationswithdi erentthresholds.Weseethatwhileinthe rstimageathresholdvalueof50wouldbeadequate,inthesecondonetheshadowsdissapearmuchearlier.Figure1:Thresholdedimageswiththresholdvaluest=80,t=50(left)andt=80,t=60(right)Itisobviousthatthereisnolowerboundfortheshadowunderneathavehicle.Butwecande neanupperlimitfortheintensityoftheshadowinordertoseparateitfromitssurround.Inthenextpartweshowhowwecande nesuchanupperboundsystematically.2.1AnUpperBoundfortheShadowAsalreadymentioneditisnoteasytodetermineanexactthresholdforanimagewhichoptimallyseparatestheshadows.Butwecandetermineanupperboundfortheirgrayvalueswhichcomesfromtheanalysisofthegrayvaluesoftheroad.Theideaisthatshadowsunderneathvehiclesarealwaysdarkerthananystainoftheroad([5]).Thusifwehavearoughestimateofthegrayleveloftheroadwecanexpectthattheshadowswillbebelowthislevel.In[6]colorinformationisusedtosegmenttheroad.In[4]theuseofentropyisproposedforthedeterminationoffreedrivingspace.Weusedanalternativesolutionthatgivesaroughapproxima-tionoffreedrivingspacebyde ningthelowestcentralhomogenousregionintheimagedelimitedbyedges.Theupperboundof