robust

EurographicsSymposiumonGeometryProcessing(2005)

n,nn(Editors)

RobustGlobalRegistration

NatashaGelfand†††HelmutPottmann‡

†ComputerGraphicsLaboratory,StanfordUniversity

‡GeometricModelingandIndustrialGeometry,ViennaUniversityofTechnology

Abstract

Weprentanalgorithmfortheautomaticalignmentoftwo3Dshapes(dataandmodel),withoutanyassumptions

orithmcomputesforeachsurfacepointadescriptorbadonlocalgeometry

numberoffeaturepointsareautomaticallypickedfromthedatashapeaccording

hfeaturepointonthedata,weuthedescriptor

valuesofthemodeltofidevelopafastbranch-and-boundalgorithm

badondistancematrixcomparisonstolecttheoptimalcorrespondencetandbringthetwoshapesinto

ultofouralignmentalgorithmisudastheinitializationtoICP(iterativeclost

point)anditsvariantsforfiorithmcanbeudformatching

shapesthatoverlaponlyoverpartsoftheirextent,forbuildingmodelsfrompartialrangescans,aswellasfor

simplesymmetrydetection,andformatchingshapesundergoingarticulatedmotion.

uctionandBackground

Globalregistration,oroptimalalignmentoftwothree-

dimensionalshapesinarbitraryinitialpositions,isafun-

damentalprobleminshapeacquisitionandshapemodel-

woshapes,oftencalledthemodelanddata,the

goalistofindarigidtransformthatoptimallypositions,or

registers,ocess

ispartofmost3Dshapeacquisitionpipelines,wherelf-

occlusionsandscannerlimitationsusuallyrequiretheac-

da

completemodel,thepartialscansneedtobebroughtinto

acommoncoordinatesystem(Figure

1),whichisusually

oblemisparticularly

hardwhennoinformationisavailableabouttheinitialposi-

tionofthemodelanddatashapes,theinputscontainnoi,

andtheshapesoverlaponlyoverpartsoftheirextent(and

theoverlapsmaynotbeknowninadvance).

Solutionstotheregistrationproblemfallintotwogen-

ss,knownasvotingmethods,makesu

ofthefactthattherigidtransformislow-dimensionaland

exhaustivelyarchesforthesmallnumberofparameters

†

ThisrearchwassupportedinpartbyNSFgrantsCARGO-

0138456andFRG-0454543,AROgrantDAAD19-03-1-033,aMax

PlanckInstitutfellowshipandaStanfordGraduateFellowship.

‡

ThisrearchwassupportedbytheAustrianScienceFundunder

grantP16002-N05.

lizedHough

transform[HB94],geometrichashing[WR97],andpo

clustering[Sto87]quantizethetransformationspaceintoa

Figure1::10in-

putscans(shownhereingoodpositionforvisualization,the

actualinputpositionsarearbitrary).Bottomleft:Registra-

tionafterapplyingouralgorithmtooverlappingpairsof

right:RegistrationafterapplyingICPander-

rorrelaxationtotheinitialpoproducedbyouralgorithm.

c