ebook img

Structure from Motion: Theoretical Foundations of a Novel Approach Using Custom Built Invariants PDF

15 Pages·0.15 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Structure from Motion: Theoretical Foundations of a Novel Approach Using Custom Built Invariants

Structure from Motion: Theoretical Foundations of a Novel Approach Using Custom Built 2 0 Invariants ∗ 0 2 n Pierre-Louis Bazin1 MireilleBoutin2 a J 1 DepartmentofEngineering, 2 2 2 DivisionofAppliedMathematics, BrownUniversity ] V ProvidenceRI02912,USA. C . s c [ Abstract 1 Werephrasetheproblemof3Dreconstructionfromimagesintermsofinter- v sectionsofprojectionsoforbitsofcustombuiltLiegroupsactions. Wethenuse 9 analgorithmicmethodbasedonmovingframesa` laFels-Olvertoobtainafunda- 1 mentalsetofinvariantsofthesegroupsactions. Theinvariantsareusedtodefine 0 a set of equations to be solved by the points of the 3D object, providing a new 1 techniqueforrecovering3Dstructurefrommotion. 0 2 0 1 Introduction / s c Theconceptofinvarianceisofmajorimportanceinmoderngeometry. Inthefieldof : v computervision,invariantshavebeenusedformorethanadecadeforobjectrecogni- i X tionandreconstruction(seeforexample[7,8]). One particular vision problem one could hope to solve using invariantsis that of r a reconstructinga3Dobjectfromasetofpicturestakenfromunknownviewpoints. For example, given n ordered points , ,... , R3 (the object) and t sets of n 1 2 n orderedpointsp ,... ,p R2 (fOoreaOchoftheOpic∈tures),wewouldliketodetermine 1 n theobjectinR3fromtheim∈ages. Onepossiblewayofsolvingthisproblemwouldbetodefineanequivalencerela- tionbetweenall the possiblepicturesof an objectandto find functionscomputedon the picturesthatare constanton eachequivalenceclass. Characteristicsof the object couldbeinferredfromthesefunctions,regardlessofthecamerapositionrelativetothe object. To define such an equivalence class, one can try to use the orbits of a group action.Realvaluedfunctionsthatareconstantontheorbitsofagroupactionarecalled ∗ThisworkwassupportedbyNSFgrantsKDIBCS-9980091and0074276. 1 invariants. We wouldthusbeinterestedin findinginvariantsofa groupactionwhich wouldbetransitiveonthesetofpicturesofanygivenobject,i.e.“viewinvariants”. Unfortunately, as is commonly known in the vision community, view invariants do not exist for 3D point sets of arbitrary size (in general position). One can still buildinvariantsforspecificobjects(forinstance,planarsetsofpoints,pencilsoflines, etc.) butnotforarbitraryshapes. Theproblemisthatthe setofpicturesofanobject intersectswith the set of picturesof other objects. Observethatif a view invariantI takesaconstantvaluecforallpicturesofanobjectO,thenI isalsoequaltoconthe setofpicturesofanyobjectwhosesetofpicturesintersectwiththesetofpicturesof O. Onecanactuallyshow[2]thatanyequivalencerelationbetweenallthepicturesof eachobjectsdefinesauniqueequivalenceclassonthespaceofpicturesandthus,any viewinvariantistrivial. Fromagrouppointofview,thismeansthatanygroupaction thatistransitiveonthesetofpicturesofanyobjectmustbetransitiveonthesetofall pictures. Another way to solve this problem would be to use the reverse approach: try to characterizeallthepossibleobjectscorrespondingtoagivenpicture. Itwouldbeuse- fultofindfunctionsthatareconstantonequivalenceclassesthatincludealltheobjects corresponding to a given picture. However, it is easy to see that such equivalence classesofobjectsareinonetoonecorrespondencewiththeequivalenceclassesofpic- turesdiscussedearlier,andthusthatonlyonesuchequivalenceclasscanexist. There are thereforeno“objectinvariants”. Nevertheless, givena view ofan object, we can inferinformationabouttheobject,sotheremustbeawaytoovercomethisdifficulty. Itisthefactthatweknowhowacamerabuildsimages(usingtheperspectivepro- jection)thatwillallowustobuildanequivalencerelationwhichcharacterizesobjects correspondingtodifferentpictures.Thekeyistoconsideranequivalencerelationona higherdimensionalspace,sortofliftingthesetofobjectsofdifferentpicturestodiffer- ent“heights”intheextradimensions.Wecandothisusingthethreeextradimensions providedbythecameracenterposition. Moreprecisely,wecanconstructaLiegroup action on the object points ,... , and the camera center P which summarize 1 n 0 O O whatisunknownabouttheobject-camerasystemgivenapictureofanobject. Invari- antsofthisgroupactionsprovetobesufficientforsolvingtheproblemofrecovering theobjectcoordinatesinR3,i.e. the”structurefrommotion”problem. Recentadvancesinthetheoryofmovingframes[3,4]provideuswithasystematic waytoproceedtoobtaintheinvariantsofanygiven(regular)Liegroupaction. Using thissystematicmethod,weobtaintheinvariantsofourcustombuiltactionwithoutany difficulty. Wecanevenventureabitfurtherbymodifyingthegroupactionandrepeat theexercisesotoobtainevenmorenewtoolsfor3Dstructureandmotionanalysis. We begin our exposition with a summary of some of the relevant theoretical as- pectsofthetheoryofmovingframesanditsapplicationtothecomputationof(joint) invariants. In Section 3, we formulate the problem of structure from motion in this framework and obtain the corresponding invariants. We end the section by remark- inghowthisleadstoa simple testtoidentifycameramotionsthatarepurerotations. Section4presentssomevariationsofourapproachwheremorecameraparametersare considered,includingageneralizationtothecaseofavariablefocaldistance. 2 2 Theoretical Foundations LetM beanm-dimensionalsmooth(Hausdorff)manifoldandGbeanr-dimensional Liegroupacting(smoothly)onM. Definition2.1. AninvariantisafunctionI :M Rwhichremainsunchangedunder → theactionofthegroup.Inotherwords, I(g z)=I(z), forallz M andallg G. · ∈ ∈ Alocalinvariantis afunctionI :M RforwhichthereexistsN ,aneighborhood e → oftheidentitye G,suchthat ∈ I(g z)=I(z), forallz M andallg N . e · ∈ ∈ Definition2.2. We say that G acts semi-regularly on M if all orbits have the same dimension. If, in addition, any point p M is surrounded by an arbitrarily small 0 ∈ neighborhoodwhoseintersectionwith theorbitthroughp isconnected,thenwesay 0 thatGactsregularly. The following theorem, due to Frobenius [5], is of central importance to our ap- proach.Aproofcanbefoundin[9]. Itprovidesuswithasimplewayofcharacterizing theorbitsusinginvariants. Theorem2.3(FrobeniusTheorem). IfGactsonanopensetO M semi-regularly ⊂ with s-dimensionalorbits, then p O there exist m s functionallyindependent 0 ∀ ∈ − localinvariantsI ,... ,I definedonaneighborhoodU ofp suchthatanyother 1 m−s 0 local invariant H defined near p is a function H = f(I ,... ,I ). If G acts 0 1 m−s regularly on O, then we can choose I ,... ,I to be global invariants on O. In 1 m−s that case, two points p ,p O are in the same orbit relative to G if and only if 1 2 ∈ I (p )=I (p ),foralli=1,... ,m s. i 1 i 2 − By functional independence of the (smooth) functions I ,... ,I on an open 1 m−s set O, we simply mean that the Jacobian matrix of I ,... ,I has maximal rank 1 m−s m s on an open and dense subset of O. The set I ,... ,I is often called a 1 m−s − { } completefundamentalsetofinvariantsonO. Notethatthecompletefundamentalset ofinvariantsisbynomeanunique. As we shallconsideractionsonpoints, we are interestedinthe case whereM = ... (n-times)=: ×(n)istheCartesianproductofncopiesofamanifold V×V× ×V V . V Definition2.4. WesaythatGactsdiagonallyon ×(n) ifthereexistsanactionofG on such that for any g G, for any n N anVd any z ,... ,z , the action 1 n V ∈ ∈ ∈ V g (z ,... ,z )canbewrittenas 1 n · g (z ,... ,z )=(g z ,... ,g z ). 1 n 1 n · ∗ ∗ The groupactions we will define for our object-camerasystems are not diagonal actions.However,wewillfindanormalsubgroupH ofGsuchthat 3 1. ThesubgroupH canbewrittenasH =H ... H(n-times)whereHisaLie × × groupactingon inaconsistentmannerwiththeactionofH,i.e.(h ,... ,h ) 1 n V · (z ,... ,z )=(h z ,... ,h z ),forallh ,... ,h Handallz ,... ,z 1 n 1 1 n n 1 n 1 n ∗ ∗ ∈ ∈ . Thisinsurethat ×(n)/H canbewrittenas ×(n)/H =( /H)×(n). V V V V 2. TheactionofG/H on ×(n)/H =( /H)×(n) isdiagonal. V V So for all practical purposes, we shall ultimately have to deal with diagonal group actions. Inourapproachtostructurefrommotion,invariantsareusedtoobtainedequations thatmustbesatisfiedbytheobjectandthecamera. Themoreinvariantswehave,the moreequationsneed to be satisfied. We need enoughequationsto completelydeter- minetheobject. Observethatthedimensionoftheorbitisboundedbythedimension of the group. So in the case of a diagonalaction, taking moreand more copiesof V (i.e. moreandmorepoints)allowsfortheexistenceofasmanyinvariantsasnecessary. The question that remains is: how can we obtain an expression for these invariants? Thanks to recent advances in the theory of moving frames [3, 4], this problem can nowbesolvedinanalgorithmicfashion. Wenowsummarizesomeofthistheoryand showhowmovingframescanbeusedasatooltoobtainacompletesetoffundamental invariants. Definition2.5. A(right)movingframeisamapρ:M Gwhichis(right)equivari- → ant,i.e.ρ(g z)=ρ(z)g−1,forallg Gandz M. · ∈ ∈ Unfortunatelymovingframesdonotexistforallgroupactions. Theorem2.6. Amovingframeexistsifandonlyiftheactionifthegroupactionsatis- fies g G z M,g z =z =e, { ∈ |∃ ∈ · } whereedenotestheidentityinG. Thispropertyiscalledfreeness ofthegroupaction. Demandingfreenessofthegroupactionisverystrong. Itappearsthat,inorderto beabletodealwiththegenericcases,weneedtorelaxthisconditionalittlebit. Definition2.7. A local moving frame is a map ρ : M G such that ρ(g z) = → · ρ(z)g−1,forallg N ,aneighborhoodoftheidentitye G,andallz M. e ∈ ∈ ∈ Theorem2.8. A localmovingframe exists ifandonlyif thereexists a neighborhood N oftheidentityine Gsuchthat e ∈ g N z M,g z =z =e, e { ∈ |∃ ∈ · } or equivalently, if and only if for all z M, the dimensionof the orbit throughz is ∈ equal to r, the dimension of G. This property is called local freeness of the group action. 4 Our hope is that, by acting diagonally on more and more copies of a manifold, weshalleventuallyobtainalocallyfreeaction. Soweareinterestedindetermininga simple criterion on the groupaction of G on to determinewhetheror notthis will V workout. Definition2.9. WesaythatGactsonM effectivelyif g G g p=p, forallp M = e . { ∈ | · ∈ } { } WesaythatGactsonM locallyeffectivelyif g G g p=p, forallp M isadiscretesubgroupofG. { ∈ | · ∈ } Many groupsdo not act effectively. However, given G acting non effectively on M,wecanconsiderG˜ = G/G ,whereG = g Gg z = z, z M ,which M M { ∈ | · ∀ ∈ } acts in essentially the same way as G except that it acts effectively. Unfortunately, effectivenessisnotquiteenoughforachievingourgoal. Definition2.10. WesaythatGactseffectivelyonsubsetsofM if,foranyopensubset U M, ⊂ g G g p=p, forallp U = e . { ∈ | · ∈ } { } WesaythatGactslocallyeffectivelyonsubsetsofM if,foranyopensubsetU M, ⊂ g G g p=p, forallp M isadiscretesubgroupofG. { ∈ | · ∈ } Observe that effectiveness on subsets implies effectiveness. The converse is not trueingeneral,butitholdsforallanalyticgroupactions. Theorem2.11. IfagroupGactsonamanifold locallyeffectivelyonsubsets,then forn Nbigenough,theinduceddiagonalactioVnofGon ×(n)islocallyfreeonon ∈ V openanddensesubsetof ×(n). Thisisequivalenttosayingthattheorbitdimension V is equal to the dimension of G on this open and dense subset. We denote by n the 0 minimalintegerforwhichthisistrue. Soatleastforallanalyticgroupactions(modingoutbyasubgroupifnecessary), acting (diagonally)on more and more copies of a manifold will eventually lead to a regular and locally free action on an open subset. A local moving frame will thus exist and provideus with the tools we need to obtain a complete fundamentalset of invariantssocompletelycharacterizetheorbits. Wenowexplainhowtoconstructa(local)amovingframeandtoobtainacomplete fundamentalsetofinvariants. Amoredetailedexpositioncanbefoundin[9,Chapter 8]. Letg = (g ,... ,g )belocalcoordinatesforGinaneighborhoodoftheidentity. 1 r SupposethatGactsregularlyonM. Forsimplicity,letusassumeinadditionthatthe orbitsofGhavethesamedimensionrasGitself.Inotherwords,weareassumingthat the action is locally free. A simple variationallowingus to dealwith merelyregular actionswillalsobeexplainedshortly. 5 Step1:Writedownthegrouptransformationequationsx¯=g xexplicitly. • · x¯ = f (g ,... ,g ,x ,... ,x ), 1 1 1 r 1 m . .  .  x¯ = f (g ,... ,g ,x ,... ,x ). m m 1 r 1 m Step 2: Choose constantsc ,... ,c R and set r of the transformedcoordi- 1 r • ∈ nates equal to those constants. For simplicity, we relabel the coordinates and write f (g ,... ,g ,x ,... ,x ) = c , 1 1 r 1 m 1 . . (1)  .  fr(g1,... ,gr,x1,... ,xm) = cr. Theseequationsarecalledthenormalizationequations. Step 3: Solve the normalizationequationsfor g = (g ,... ,g ). The solution 1 r • g =ρ(x)isamovingframe. Step4: Computetheactionofthemovingframeontheremainingcoordinates. • Thesetofresultingfunctions x¯ = I (x ,... ,x ), r+1|g=ρ(x) 1 1 m . .  .  x¯ = I (x ,... ,x ). m|g=ρ(x) m−s 1 m isacompletefundamentalsetoflocalinvariants. ThechoiceofconstantsinStep2issomewhatarbitrary:wearefreetochooseany numbersforwhichasolutiontothenormalizationequationsexists,providedthatthese constantsdefinea cross-section(i.e.providedthatthe normalizationequationsdefine asubmanifoldwhichistransversaltotheorbits). Tosimplifythesolvingprocess,itis usuallyagoodideatochooseasmanyconstantsaspossibletobezero. Iftheactionisnotfreebutmerelyregular,wecanstillfindasystemoffunctionally independentlocalinvariants. Whatwedoisthefollowing. Letsbethedimensionof the orbitsof G (s < r). We solve the s equationsf (g,x) = c ,... ,f (g,x) = c 1 1 s s for s of the group parameters and replace them in the remaining equations x = s+1 f (g,x),... ,x = f (g,x)togetthem sinvariants. Theothergroupparame- s+1 m m − terswillnotappearintheequations. Thisprocedureiscalledapartialmovingframe normalizationmethod. Equipped with these theoretical tools, the computation of invariants becomes a simple systematic procedure. We can thusfeel free to considerany Lie groupaction imaginable and try to obtain its invariants. As we have seen, in theory, results are guaranteedprovidedthatthegroupactionislocallyeffectiveonsubsets,whichwecan alwaysarrangeinthecaseofanalyticgroupactions. 6 3 Application to 3D Shape Reconstruction Letusthinkforamomentabouttheprocessoftakingapicture.Thisprocessinvolves, firstof all, the placementof a camerain space. Then, particlesoflightstarting from theshape(pointsinR3)travelonastraightlineinthedirectionofthecameracenter, leavingitstraceonafilm, i.e.ontheintersectionofthepictureplaneandthestraight travelline.Sotothepicture-camerasystemplacedsomewhereinR3,therecorresponds asetofn straightlinesinR3 representingpathsoflightgoingfromthe objectto the cameracenter. ThisprocesscanbeseenasagroupactiongeneratedbyanactionofSE(3)andan actionofRnonthecameracentertogetherP withtheimagepointsP ,... ,P . The 0 1 n ideaistoalloweachP tomoveindependentlyalongthelinethroughP andP ,while i i 0 allowingthelineconfigurationtoberotatedandtranslatedinspace.Asdiscussedinthe introduction,includingthecameracenteronthespaceactedonbythegroupwillallow ustoobtainsignificantequivalenceclasses. Thisisthekeytoguaranteetheexistence ofnon-trivialinvariantsandtheseinvariantscanbeobtainedbyoursystematicmethod. So, given is a 2D image depicting n points p ,... ,p R2. We assume this 1 n ∈ picturewastakenbyacamerawithfixedinternalparameters.Theseparameterscanbe calibratedbeforehand,sothatthefocallengthis =11andthe2Dimagecoordinates F matchthe 3D coordinatesasdefinedbelow. We embedthe picture-camerasystem in R3 by setting the camera center to be p˜ = (0,0,0) and the picture points p˜’s to 0 i be p˜ = p . This is of course, in general, not the actual position in which the i i ×F picturewastaken. However,thereexistsarigidtransformationg SE(3)suchthat ∈ g (p˜ ,p˜ ,... ,p˜ ) = (P ,P ,... ,P ) corresponds to the actual position of the 0 1 n 0 1 n · picture-camerasystematthemomentwherethepicturewastaken. Inotherwords,if theobjectismadeofnpoints,say ,... , ,theneachtransformedimagepointP 1 n i O O liesonthestraightlinepassingthrough andthecameracenterP . Inordertofully i 0 O formulatetheproblemintermsoforbits,wewanttoconsidersmoothtransformations thatwillmaptheimagepoints(P ,... ,P )totheobjectpoints( ,... , ). For 1 n 1 n O O this, we allow each pointP to moveindependentlyalong each ray of light so to go i backtoitssourceontheobject. We would like to determine where P and the ’s lie. Given a picture, it is of 0 i O course impossible to determine the points (P , ,... , ). However the “line ar- 0 1 n O O rangement”definedbythepicture-camerasystemprovidesuswithimportantinforma- tion. In particular, we know that the points (P , ,... , ) lie on the orbitof the 0 1 n Liegroupactionon (P ,P ,... ,P ) R3 (RO3)×(n) dOefinedby: 0 1 n { ∈ × } P¯ = RP +T 0 0 P¯ = R(P +λ (P P ))+T, fori=1,... ,n, i i i i 0 − with R SO(3) a rotation, T R3 a translation and λ R, a factor of depth, i fori = 1∈,... ,n. Observethatth∈e actionof Rn parameterize∈dbythe λ’s commutes withtheactionofSE(2)generatedbytherotationRandthetranslationT.Therefore, 1thevalueisarbitrary,itsimplyfixestheoverallscaleofthe3Dreconstruction. 7 thisdefinesafinitedimensionalLiegroupaction,morepreciselytheactionofa(6+ n)-dimensionalLie groupon a (3n+3)-dimensionalmanifold. Note that, although invariant-basedtechniqueshavealreadybeenusedinthefieldforthegeneralprojective andaffinetransformationgroups[11,13,12],theLiegroupsweshalldefinehavelittle todowiththesetraditionaltransformationgroups. Assumingthatthepicturepointsaredistinct, thenthegroupactionisregularand the orbitsare 6-dimensional,for n = 1, and(6+n)-dimensionalas soonas n 2. ≥ ThereforebyTheorem2.3,thereare2n 3fundamentalinvariantswhenevern 2. − ≥ Wenowfollowthestepsofthemovingframenormalizationmethodtoobtainthem. Weset P¯ = (0,0,0)T, 0 P¯ (0,1,0) = 0, 1 · P¯ (0,0,1) = 0, 1 · P¯ (0,0,1) = 0, 2 · andP¯ (1,0,0) = 1, foralli=1,... ,n. i · Solvingforthegroupparameters,weobtain T = RP , 0 − R = R R R , 1 2 3 1 0 0 0 f g R1 =  √f2+g2 √f2+g2 , 0 g f  −√f2+g2 √f2+g2     √x21+y12 0 z1  √x21+y12+z12 √x21+y12+z12 (2) R =  0 1 0 , 2  −z1 0 √x21+y12   √x21+y12+z12 √x21+y12+z12   x1 y1 0  √x21+y12 √x21+y12 R = y1 x1 0 , 3  −√x21+y12 √x21+y12  0 0 1   λ =  1 1.  i (R(Pi−P0))x − wheref = −y1x2+x1y2, g = z2(x21+y12)−z1(x1x2+y1y2), (x ,y ,z )T = P P and √x21+y12 √x21+y12√x21+y12+z12 1 1 1 1 − 0 (x ,y ,z )T = P P . Thesegroupparametersdefinea movingframe(MF).Re- 2 2 2 2 0 − 8 placingthemovingframeintothetransformationequations,weget: P¯ = (0,0,0)T, 0 MF P¯1(cid:12)(cid:12)MF = (1,0,0)T, (cid:12) 1 P¯ (cid:12) = f√x21+y12+z12(x1y2−x2y1)+g[z2(x21+y12)−z1(x1x2+y1y2)] , 2 MF  (x1x2+y1y2+z1z2)√x21+y12√f2+g2  (cid:12)(cid:12)  0    1 f√x21+y12+z12(x1yi−xiy1)+g[zi(x21+y12)−z1(x1xi+y1yi)] P¯i MF =  (x1xi+y1yi+z1zi)√x21+y12√f2+g2 . (cid:12)(cid:12)  g√x21+y12+(xz121x(xi+iyy11−yxi+1yzi1)z+i)f√[zix(21x+21+y12y√12)f−2z+1(gx21xi+y1yi)]    foralli=3,... ,n,where(x ,y ,z )=P P . Eachcomponentofthesevectorsis i i i i 0 − aninvariantofthegroupaction. Letustrytounderstandthegeometricmeaningoftheseexpression. Observethat f2+g2 = k(x1,y1,z1)×(x2,y2,z2)k. After a few manipulations, we can rewrite the √x21+y12+z12 pabovesystemas: P¯ = (0,0,0)T, 0 MF P¯1(cid:12)(cid:12)MF = (1,0,0)T (cid:12) 1 P¯ (cid:12) = k(x1,y1,z1)×(x2,y2,z2)k , 2 MF  (x1,y1,z1)·(x2,y2,z2)  0 (cid:12) (cid:12)   1 P¯ = [(x1,y1,z1)×(xi,yi,zi)]·[(x1,y1,z1)×(x2,y2,z2)] . i MF  [(x1,y1,z1)·(xi,yi,zi)]k(x1,y1,z1)×(x2,y2,z2)k  (xi,yi,zi)·[(x2,y2,z2)×(x1,y1,z1)]k(x1,y1,z1)k (cid:12)(cid:12)  [(x1,y1,z1)·(xi,yi,zi)]k(x1,y1,z1)×(x2,y2,z2)k    Itnowbecomesclearerthatthecomponentsof P¯ and P¯ aresineorco- 2 MF i MF sineofanglesbetweenthedirectionsspannedbyP P ,P P ,P P andthedirections 1 0(cid:12) 2 0 i (cid:12)0 orthogonal to them. These are clearly invariant by tr(cid:12)anslation, ro(cid:12)tation, and motion along the projection lines. As a fundamental set, we simply pick the only 2n 3 − non-constantinvariants: (x ,y ,z ) (x ,y ,z ) 1 1 1 2 2 2 I = k × k 2 (x ,y ,z ) (x ,y ,z ) 1 1 1 2 2 2 · [(x ,y ,z ) (x ,y ,z )] [(x ,y ,z ) (x ,y ,z )] 1 1 1 i i i 1 1 1 2 2 2 I = × · × i [(x ,y ,z ) (x ,y ,z )] (x ,y ,z ) (x ,y ,z ) 1 1 1 i i i 1 1 1 2 2 2 · k × k (x ,y ,z ) [(x ,y ,z ) (x ,y ,z )] (x ,y ,z ) i i i 2 2 2 1 1 1 1 1 1 J = · × k k i [(x ,y ,z ) (x ,y ,z )] (x ,y ,z ) (x ,y ,z ) 1 1 1 i i i 1 1 1 2 2 2 · k × k 9 fori=3,... ,n. Each picturetaken definesa pointin R3 (R3)×(n) andthereforedeterminesan × orbitofourgroupaction.Eachorbitischaracterizedbythesetof2n 3equationsgiven − by the invariants. More precisely, indexing the pictures with the discrete parameter τ =1,... ,t,wehave I (Pτ,P ,... ,P ) = ατ, fori=2,... ,n, i 0 1 n i J (Pτ,P ,... ,P ) = βτ, forj =3,... ,n. j 0 1 n j forappropriateconstantsατ’sandβτ’s. Theseconstantsareprescribedbythepictures: i j sincethepicture-camerasystemitselfbelongstotheorbits,wehave ατ = I (p˜τ,p˜τ,... ,p˜τ) i i 0 1 n βτ = J (p˜τ,p˜τ,... ,p˜τ) j j 0 1 n Weareinterestedinsolvingtheequations I (Pτ, ,... , ) = ατ, fori=2,... ,n i 0 O1 On i Jτ(Pτ, ,... , ) = βτ, forj =3,... ,n. j 0 O1 On j forτ =1,... ,t. For Wehave(2n 3)t(non-linear)equationswith3n+3tunknowns,thesolutionof − whichisdetermineduptoarotationandtranslationofthe3Dcamera-objectsystemas awhole,whichcancanfixarbitrarily,thuseliminatingsixvariables2. Forn > 3and t 3n−6, the numberof equationisgreaterthanthenumberofunknownso we can ≥ 2n−6 trytosolvethem. Experiments with real video images have been performed (see Fig.1 using a se- quential non-linear optimization technique based on the Levenberg-Marquardtalgo- rithm[10].Thefeaturepointsusedinthepicturesareendpointsoflinesandrectangles associatedfromoneimagetothenextwithatrackingprocedure[1].Thereconstructed 3Dobjectisvalidinanyview,evenifthebottomandleftsideelementsarenotperfectly replaced,duetothetrackingnoise. Thecomputationstakeonlyafewminutes. Observethatourcamera-systemdoesnottakeintoaccounttheangleofthecamera; theorientationoftheimageplanewasnotusedinourdescriptionofthecamera-picture system, Our invariants are invariant under a rotation of the image plane and so we cannot use them to recover the camera orientation. But this has its advantages, as illustratedbythefollowinglemma. Lemma3.1. Themotionofthecamerabetweentwopicturesisapurerotation(i.e.a rotationaroundthecenterofprojectionP )ifandonlyifthevaluesoftheinvariants 0 I ,J i=2,... ,n,j =3,... ,n evaluatedonanycorrespondingpointsinthetwo i j { | } viewsareequal. 2However, we should keep in mind that the choice of these variables will affect the numerical resolution[6]. 10

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.