搜索
ResearchProposalN.MavaddatDepartmentofComputerScienceandSoftwareEngineeringTheUniversityofWesternAustraliaCrawley,W.A.6009navid@cs.uwa.edu.auApril2002A.ProposedStudy1.TitleAutomaticDeterminationofVanishingPointsandGeometricCuesforSingleViewRectificationandMetrology.2.ContributionFromasingletwo-dimensionalimageofathree-dimensionalobjectorsceneitispossibletoextractsubstantialinformationabouttheoriginalobjectorevenpartiallyreconstructtheimageoftheobjectfromadifferentviewingangle.[4]Thishasmanyapplicationsinthefieldsofarchitecture,archaeology,forensicscienceandmanyothers.Thisanalysiscanprovide,inthecaseofmetrology,measurementoftheoriginalscene,andinthecaseofrectification,thenecessaryinformationtoconstructavirtualmodelofthescene.Acriticalpartoftheanalysisprocessistoaccuratelyidentifyattributesofinterest.Onesuchattributeofinterestarethe‘vanishingpoints’ofascene.Thesearedefinedaspointsatwhichtheextensionsofparallellinesappeartoconvergeintheperspectiveviewoftheimage.Inpracticethedetectionofthevanishingpointsismadebyvisualinspectionoftheimageandisbothcumbersomeandnotveryaccurate.Anautomaticmethodofthevanishingpointdetectionwouldeliminatethesubjectiveinspectionoftheimageandexpeditetheprocess.InmyHonoursprojectIwasabletoadvanceamethodforautomaticdetectionofthevanish-ingpointsthroughthecombinationofimageanalysistechniquesandprojectivegeometry.Themethodwassuccessfullyappliedtothedetectionofthevanishingpointsforseveralimagesandreconstructionofimagesfromdifferentperspectives.However,theabovemethodcanonlybereadilyimplementedwhentherearesufficientstraightlinesintheimagethatcanbedetectedbycomputationalmethodsandtheirconvergencepointdetermined.Ingeneralthetwo-dimensionalimagesmaycontainnoneoraveryfewstraightlines.Hencetheproposedmethodhastobemuchimprovedandaugmentedtobeausefultoolforpracticalapplications.Theattributesofinterestarenotconfinedtovanishingpoints.Otherattributessuchahorizonlinesarealsousefulforsingleviewimageanalysis.Theseattributeshaveyettobeexploredandtheirdetectionprocessesautomated.Furtherworkincludesthedevelopmentofprototypesforproposedapplicationsofsingleviewmetrologyandrectification.Inadditiontovanishingpoints,othergeometricalcuescangreatlyassisttheaccuracyandfeasi-bilityofsingleviewmetrology.Thesegeometricalcuesincludedeterminationofrightanglesandcirclesinascene.B.ResearchPlan1.TimeEstimatesforCompletionDateActivity05/2002Literaturereviewandresearchproposal.08/2002Comparisonofcurrentvanishingpointdetectorswithrespecttoefficiencyandaccuracy10/2002Extensionofexistingtechniquesbeyondstraightlinestoincludehorizondetection11/2002Investigationofdetectionofknownangles.02/2003Investigationofdetectionofknowncurveshapes.05/2003Analysisoftextures.01/2004Erroranalysisoftechniquesinvestigated04/2004ThesisOutline.05/2004ThesisComposition.10/2004ThesisReview.11/2004FinalSubmission.2.ProjectAims²Replicateprocessofsingleviewrectificationandmetrologywithouttheneedofcameracalibration.²Improveviewreconstructionbytheanalysisofgeometriccues.²Compareaccuracyandefficiencyofvanishingpointdetectionalgorithms.²Investigatetheaccuracyofgeometricfeaturedetectionwithrelationtohumaninteractioninthedecisionmakingscheme.²Investigatethesignificanceoflesscommonlyusedgeometricfeatureintheprocessofimagerectification.3.ExistingApproaches3.1.ManualVanishingPointDetectionManualvanishingpointdetectioninvolvesvisuallyisolatingpairsoflinesimagethatareknowntoparallelinthescene,andinterpolatingtheircrossingpoints.Thesecrossingpointsdefinethevanishingpoints.Animageofathreedimensionalscenehasaverticalvanishingpointandeachplanewithahorizontalcomponentdefinesavanishingpoint,whichwilllieonthehorizonline.Linesthatareknowntobeparallelinthesceneandthatarealsoparallelin2dimageofthescenehavevanishingpointsthatlieinfinitelyfaraway.Onlythedirectionofthesevanishingpointscanbecomputed.Thisapproachrequiresmanualinputandthereforeitsaccuracyisdependantontheaccuracyofthemeasurementofparallellines.Vanishingpointsareoftenfarofftheimageandhenceaslightinaccuracyinthemeasurementofalinewillresultinalargererrorinthepositionofthecalculatedvanishingpoint.3.2.VanishingPointDetectionusingHoughTransformWhenlinesthatareparallelinworldcoordinatesaremappedthroughaprojectivetransformtoanimageplane,theywillnolongerbeparallelunlessthevanishingpointisatinfinity.Theselinewillinfactcrossandthiscrossingpointdefinesthevanishingpoint.TheHoughTransformisawidelyusedalgorithmfordetermininglinesandothergeometricalprimitivesincomputervision.Forthepurposesofvanishingpointdetection,theHoughTransformisoptimisedinordertodetectlinesthatwouldbeparallelintheworldcoordinatesofascene.Usingtheseparallellines,thevanishingpointoftheimagecanbecalculated[22]3.3.GaussianSphereManyvanishingpointdetectionalgorithmsrequiredthesearchingofmanypossiblevanishingpointcandidates.Vanishingpointsandotherpossiblepointsofinterestoftenliefaroutsidetheimagearea.Tosearchthroughthesethetraditionalwayisatimeconsumingandcumbersometaskasitisimportanttotesteachcandidaterelativetotheothers.Thismeansthatthespacetobesearchedisnotbounded.Thetimetakentoevaluat