An_Iterative_Image_Registration_Technique_with_an_Application_to_Stereo_Vision

《An_Iterative_Image_Registration_Technique_with_an_Application_to_Stereo_Vision》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、FromProceedingsofImagingUnderstandingWorkshop,pp.121-130(1981).AnIterativeImageRegistrationTechniquewithanApplicationtoStereoVisionBruceD.LucasTakeoKanadeComputerScienceDepartmentCarnegie-MellonUniversityPittsburgh,Pennsylvania15213Abstract2.TheregistrationproblemImageregistrationfi

2、ndsavarietyofapplicationsinThetranslationalimageregistrationproblemcanbecomputervision.Unfortunately,traditionalimageregistrationcharacterizedasfollows:WearegivenfunctionsF(x)andtechniquestendtobecostly.WepresentanewimageG(x)whichgivetherespectivepixelvaluesateachlocationregistratio

3、ntechniquethatmakesuseofthespatialintensityxintwoimages,wherexisavector.WewishtofindthegradientoftheimagestofindagoodmatchusingatypeofdisparityvectorhwhichminimizessomemeasureoftheNewton-Raphsoniteration.OurtechniqueisfasterbecausedifferencebetweenF(x+h)andG(x),forxinsomeregionofite

4、xaminesfarfewerpotentialmatchesbetweentheimagesinterestR.(Seefigure1).thanexistingtechniques.Furthermore,thisregistrationtechniquecanbegeneralizedtohandlerotation,scalingandshearing.Weshowshowourtechniquecanbeadaptedforuseinastereovisionsystem.1.IntroductionImageregistrationfindsava

5、rietyofapplicationsincomputervision,suchasimagematchingforstereovision,patternrecognition,andmotionanalysis.Untortunately,Figure1:Theimageregistrationproblemexistingtechniquesforimageregistrationtendtobecostly.Moreover,theygenerallyfailtodealwithrotationorotherTypicalmeasuresofthedi

6、fferencebetweenF(x+h)anddistortionsoftheimages.G(x)are:Inthispaperwepresentanewimageregistration•L1norm=åe

7、(FxhGx+-)()

8、techniquethatusesspatialintensitygradientinformationtoxRdirectthesearchforthepositionthatyieldsthebestmatch.12/Bytakingmoreinformationabouttheimagesintoaccount,•Lno

9、rm=()[]FxhGx()(+-)22åxRethistechniqueisabletofindthebestmatchbetweentwoimageswithfarfewercomparisonsofimagesthan•negativeofnormalizedcorrelationtechniqueswhichexaminethepossiblepositionsofregistrationinsomefixedorder.Ourtechniquetakesadvantageofthefactthatinmanyapplicationsthetwo=-+

10、åxReFxhGx()()images