efficientransac_Efficient RANSAC for Point-Cloud Shape Detection

Volume0(1981),Number0pp.1–12

Ef cientRANSACforPoint-CloudShapeDetection

RuwenSchnabel

RolandWahl

ReinhardKlein

UniversitätBonn,ComputerGraphicsGroup

Abstract

Inthisworkwepresentanautomaticalgorithmtodetectbasicshapesinunorganizedpointclouds.Thealgorithmdecomposesthepointcloudintoaconcise,hybridstructureofinherentshapesandasetofremainingpoints.Eachdetectedshapeservesasaproxyforasetofcorrespondingpoints.Ourmethodisbasedonrandomsamplinganddetectsplanes,spheres,cylinders,conesandtori.Formodelswithsurfacescomposedofthesebasicshapesonly,e.g.CADmodels,weautomaticallyobtainarepresentationsolelyconsistingofshapeproxies.Wedemonstratethatthealgorithmisrobusteveninthepresenceofmanyoutliersandahighdegreeofnoise.Theproposedmethodscaleswellwithrespecttothesizeoftheinputpointcloudandthenumberandsizeoftheshapeswithinthedata.Evenpointsetswithseveralmillionsofsamplesarerobustlydecomposedwithinlessthanaminute.Moreoverthealgorithmisconceptuallysimpleandeasytoimplement.Applicationareasincludemeasurementofphysicalparameters,scanregistration,surfacecompression,hybridrendering,shapeclassi cation,meshing,simpli cation,approximationandreverseengineering.

CategoriesandSubjectDescriptors(accordingtoACMCCS):I.4.8[ImageProcessingandComputerVision]:SceneAnalysisShape;SurfaceFitting;I.3.5[ComputerGraphics]:ComputationalGeometryandObjectModelingCurve,surface,solid,andobjectrepresentations

1.Introduction

Duetotheincreasingsizeandcomplexityofgeometricdatasetsthereisanever-growingdemandforconciseandmean-ingfulabstractionsofthisdata.Especiallywhendealingwithdigitizedgeometry,e.g.acquiredwithalaserscanner,nohandlesformodi cationofthedataareavailabletotheuserotherthanthedigitizedpointsthemselves.However,inor-dertobeabletomakeuseofthedataeffectively,therawdigitizeddatahastobeenrichedwithabstractionsandpos-siblysemanticinformation,providingtheuserwithhigher-levelinteractionpossibilities.Onlysuchhandlescanpro-videtheinteractionrequiredforinvolvededitingprocesses,suchasdeleting,movingorresizingcertainpartsandhencecanmakethedatamorereadilyusableformodelingpur-poses.Ofcourse,traditionalreverseengineeringapproachescanprovidesomeoftheabstractionsthatweseek,butusu-allyreverseengineeringfocuseson ndingareconstructionoftheunderlyinggeometryandtypicallyinvolvesquitete-dioususerinteraction.Thisisnotjusti edinasettingwhere

acompleteanddetailedreconstructionisnotrequiredatall,orshalltakeplaceonlyaftersomebasiceditingoperationshavebeenappliedtothedata.Ontheotherhand,detectinginstancesofasetofprimitivegeometricshapesinthepointsampleddataisameanstoquicklyderivehigherlevelsofab-straction.ForexampleinFig.1patchesofprimitiveshapesprovideacoarseapproximationofthegeometrythatcouldbeusedtocompressthepoint-cloudveryeffectively.Anotherproblemarisingwhendealingwithdigitizedgeom-etryistheoftenhugesizeofthedatasets.Thereforetheef ciencyofalgorithmsinferringabstractionsofthedataisofutmostimportance,especiallyininteractivesettings.Thus,inthispaperwefocusespeciallyon ndinganef -cientalgorithmforpoint-cloudshapedetection,inordertobeabletodealevenwithlargepoint-clouds.OurworkisahighperformanceRANSAC[FB81]algorithmthatiscapa-bletoextractavarietyofdifferenttypesofprimitiveshapes,whileretainingsuchfavorablepropertiesoftheRANSACparadigmasrobustness,generalityandsimplicity.Attheheartofouralgorithmareanovel,hierarchicallystructuredsamplingstrategyforcandidateshapegenerationaswellasanovel,lazycostfunctionevaluationscheme,whichsignif-

e-mail:{schnabel,wahl,rk}@cs.uni-bonn.de

cTheEurographicsAssociationandBlackwellPublishing2007.PublishedbyBlackwell

Publishing,9600GarsingtonRoad,OxfordOX42DQ,UKand350MainStreet,Malden,MA02148,USA.