Augmented Reality Scouting for Interactive 3D Reconstruction(2)

Sincetheinputimagescontaintoomuchredundantinformationforactualthereconstruction,themostrelevantinformationrequiredfor ndingcorrespondencesmustbeextractedbyusingfeaturepoints.Featureextractionselectsimagepointsorregionswhichgivesignif-icantstructuralinformationtobeidenti edinotherimagesshow-ingthesameobjectsofinterest.WeuseHarriscornersasfeaturepoints[5]whicharewellsuitedforsparsecorrespondencesearcheswhichisthecaseforurbanscenes.

3.3CorrespondenceandPoseEstimation

Inordertorelateasetofimagesgeometricallyitisnecessaryto ndcorrespondences.Forthetaskofcalculatingtherelativeorientationbetweenimagesitissuitabletoextractfeatureswithgoodpointlocalizationasprovidedbythefeatureextractionstep(seeabove).Therelativeorientationbetweentwoviewstakenfromcalibratedcamerascanbecalculatedfrom vepointcorrespondences.HenceaRANSAC-basedapproachisusedforrobustinitialestimationoftherelativeposebetweenthe rsttwoviews.Weutilizeanef cientprocedureforrelativeposeestimation[11]inordertotestmanysamplesquickly.Theresultofthisprocedureistherelativeorien-tationbetweenthesetwoviews,butwithunknownoverallscale.Therelativeposetranslatesintoknownepipolargeometry,whichrepresentstherelationshipofpixelsinoneviewwithimagesofthecorrespondingcameraraysinthesecondview.Withtheknowledgeoftherelativeposesbetweentwoviewsandcorrespondingpointfeaturesvisibleinatleast3images,theorientationsofallviewsinthesequencecanbeupgradedtoacommoncoordinatesystem

Figure3:Thereconstructionpipelineconsistsoffourmaincomponents:featureextraction,correspondencesearch,cameraposeestimation,anddensematching.Eachcapturedimageispassedthroughthispipelineinordertogenerateorenhancethe3Dmodel.

byusinganabsoluteposeprocedure[4],againcombinedwithaRANSACapproachtoincreasetherobustness.Theposeofeveryadditionalincomingimageiscalculatedonthebasisof2Dto3Dpointcorrespondences,whichisstrongerthanusingtheepipolarrelationshipbetweentwoviewsalone.

Purelyimagebasedreconstructionsarelocatedinalocalcoordi-natesystem,whichisupgradedtoaworld-referencesystemusingthemeasuredGPSlocationsofthecamerapositions.2Thetransfor-mationfromthelocaltotheglobalsystemisasimilaritytransforminthecaseofcalibratedcameras.

Thecameraposesandthesparsereconstructionconsistingof3Dpointstriangulatedfrompointcorrespondencesarere nedus-ingasimplebutef cientimplementationofsparsebundleadjust-ment[10].Ourimplementationallowsthere nementofthecameraintrinsicparametersandtheintegrationofGPSdatawithestimateduncertaintiesaswell.Theoutputofthisstepareoptimizedcameraorientationsandintrinsicparametersinthe rstplace.Addition-ally,sparse3Dpointscorrespondingtotheimagefeaturesvisibleinseveralviewsarere ned,too.3.4DenseDepthEstimation

Withtheknowledgeofthecameraparametersandtherelativeposesbetweenthesourceviewsdensecorrespondencesforallpixelsofaparticularkeyviewcanbeestimated.Sincetherelativeposebe-tweentheincorporatedviewsisalreadyknown,thisprocedureisbasicallyaone-dimensionalsearchalongthedepthraysforeverypixel.Triangulationofthesecorrespondencesresultsinadense3Dmodel,whichre ectsthetruesurfacegeometryofthecapturedob-jectinidealsettings.

WeutilizeaGPU-acceleratedplane-sweepapproachtogeneratethedepthmapforeachsourceview[17,18].Planesweeptech-niquesincomputervisionaresimpleandelegantapproachestoim-agebasedreconstructionwithmultipleviews,sincearecti cationprocedureasneededinmanytraditionalcomputationalstereometh-odsisnotrequired.The3Dspaceisiterativelytraversedbyparallelplanes,whichareusuallyalignedwithaparticularkeyview(Fig-ure4).Theplaneatacertaindepthfromthekeyviewinduceshomographiesforallotherviews,thusthereferenceimagescanbemappedontothisplaneeasily.

Iftheplaneatacertaindepthpassesexactlythroughthesur-faceoftheobjecttobereconstructed,thecolorvaluesfromthekeyimageandfromthemappedreferencesimagesshouldcoincideatappropriatepositions(assumingconstantbrightnessconditions).

theGPSantennaandtheprojectioncenterofthecameraare

very

close,weignoretheresultingoffsetbetweenthem.

2Since

Figure4:Planesweepingprinciple.Fordifferentdepthsthehomog-raphybetweenthereferenceplaneandthereferenceviewisvarying.Consequently,theprojectedimageofthereferenceviewchangeswiththedepthaccordingtotheepipolargeometry.

Hence,itisreasonabletoassignthebestmatchingdepthvalue(ac-cordingtosomeimagecorrelationmeasure)tothepixelsofthekeyview.Bysweepingtheplanethroughthe3Dspace(byvary-ingtheplanesdepthwrt.thekeyview)a3Dvolumecanbe lledwithimagecorrelationvaluessimilartothedisparityspaceimage(DSI)intraditionalstereo.Thereforethedensedepthmapcanbeextractedusingglobaloptimizationmethods,ifdepthcontinuityoranyotherconstraintonthedepthmapisrequired.Weemployasimplewinner-takes-allstrategytoassignthe naldepthvaluesforperformancereasons.3.5Output

Adepthmapisgeneratedasdescribedaboveforeverytripletofad-jacentviews,andthesingledepthimagesneedtobefusedintoonecommonmodel.Currently,weemployaverysimpletechnique:thedepthmapsareconvertedintocoloredpointclouds(usingthereferenceviewfortexturing),andthesepointsetsareconcatenatedtoobtainthecombinedmodel.Thisapproachallowsaneasyin-crementalupdateofthedisplayedmodelaftergenerationofanewdepthmap.Futureworkwilladdressthecreationof3Dsurfacemeshesfromthedepthmaps(e.g.[18]),whichrequiresmorecom-plexmethodstoassignatexturetotheresultingmodel.4

RESULTS

The rstprototypewastestedwithmultiplebuildingsatourcam-pus.Additionally,weusedittoreconstructanancientbrickwall

221

showninFigure5.Onlysomesmallcluttercanbeobservedinthebottomoftheresultingmodel.

Thereconstructiontimedependsontheimageresolutionandthenumberofextractedfeatures.Fortheaboveexample,eachpassofthepipelinetakeslessthanoneminuteforuploading,featureextraction, ndingcorrespondences,densematching,andupdatingthe3D

model.

Figure5:Thesescreenshotsshowatestdatasetofanoldbrickwall(with6inputimages).Theimageinthemiddleshowsthereconstruc-tionofthecamerapositionsincludingsparsepoints.Theimageonthebuttonshowsascreenshotofthe nal3Dmodel.

5CONC …… 此处隐藏：5983字，全部文档内容请下载后查看。喜欢就下载吧 ……