ebook img

Geometric, Kinematic and Radiometric Aspects of Image-Based Measurements PDF

30 Pages·2002·1.7 MB·English
by  LiuTiansh
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 Geometric, Kinematic and Radiometric Aspects of Image-Based Measurements

AIAA 2002-3239 Geometric, Kinematic and Radiometric Aspects of Image-Based Measurements Tianshu Liu NASA Langley Research Center Hampton, VA 23681-2199 22nd AIAA Aerodynamic Measurement Technology and Ground Testing Conference 24-26 June 2002/St. Louis, Missouri I For permission to cop'_' or reput;iist_, contact t_e American Institute ol Aeronautics and Astronautics t801 Aie×_nder Bell Drive. SuiteS00 Reston, VA 20191 AIAA 2002-3239 Geometric, Kinematic and Radiometric Aspects of Image-Based Measurements Tianshu Liu _ NASA Langley Research Center, Hampton, VA 23681-0001 computer vision, and other scientific and engineering Abstract disciplincs have developed various methods that are best This paper discusses theoretical foundations of suitable to particular applications in their fields. In quantitative image-based measurements for extracting and particular, both photogrammetrists and computer vision reconstructing geometric, kinematic and dynamic scientists have studied image-based techniques for many properties of observed objects. New results are obtained years to obtain metric and geometric inlormation. The by using a combination of methods in perspective approaches developed by photogrammetrists arc more geometry, differential geometry, radiometry, kinematics mature and quantitative, which are recently extended to and dynamics. Specific topics include perspective non-topographic applications [1]. By contrast, in order to prqiection transformation, perspective developable conical deal with more complicated vision problems relatcd to surface, perspective projection under surface constraint, artificial intelligence, computer scientists tend to adopt perspective invariants, the point correspondence problem, morc versatile mathematical approaches in perspective motion fields of curves and surfaces, and motion equations geometry, differential geometry and image algebra [2-5]. of image intensity. The methods given in this paper arc However, the approaches used by computer vision useful for determining morphology and motion fields of scientists are of qualitative nature in many cases and detormable bodies such as elastic bodies, viscoelastic generally less accurate than those used in photogrammetry mediums and fluids. in metric measurements. Because the objectives of Table of Contents different disciplines are very different, there is a lack of sufficient interaction among specialists in various technical 1.Introduction communities. Perhaps due to different notations, jargons 2. Perspective Prqiection Transformation from 3D Space and methodologies in these communities, it is difficult to to 2D Image transcend the different technical domains and see a unified 3. Projective Developable Conical Surlace Containing 3D scope of various image techniques. Curve From a methodological standpoint, the approaches in 4. Perspective Projection under Surthce Constraint photogrammetry and computer vision should be integrated 5. Perspective Projection of Motion Field Constrained on into a universal theoretical framework. Furthermore, Surface unlike computer vision scientists who mainly study rigid 6. The Correspondence Problem bodics, aerospace engineers and scientists often deal with 7. Composite Image Space and Object Space complex morphology and motion fields of deformable 8. Perspective Invariants of 3D Curve bodies such as elastic bodies, viscoelastic mediums and 9. Modeling of Imaging System fluids. It is highly desirable to formulate universal 10. Typical Radiation Processes theoretical foundations lor quantitative image-based I1.Reflection and Shape Recovery measurements of morphology and motion fields of 12.Motion Equations of Image Intensity deformable bodies. In this paper, we will focus on the 13.Conclusions geometric, kinematic and radiometric aspects of image- based measurements. First, we will provide a unified 1.Introduction treatment of the perspective projection transtbrmation Image-based measurement techniques play an from the 3D object space to the 2D image plane and increasingly important role in virtually all natural sciences illustrate geometric connections among different and engineering disciplines since they can provide formulations of the perspective proiection translormation. tremendous information and knowledge about observed Then, we will discusses some specific problems for objects in a global, non-contact way with high temporal recovering geometry and motion, such as projective and spatial resolution. Specialists in photogrammetry, developable conical surface, projection under surfacc constraint, reconstruction of motion field on a surface and motion field of a 3D curve, the correspondencc problem, Research Scientist. Model Systems Branch. MS 238, Member AIAA Copyright © 2002 by the American Institute of Aeronautics and and projective invariants. This is an area tor combined Astrorlautics, Inc. No copyright is asserted in the United States under application of approaches in perspective geometry, Title 17. U.S. Code. The U,S. Govemment has a royalty-free licensc to differential geometry, kinematics and dynamics. In the exercise all rights under the copyright claimed herein for Governmental purposes. All other rights are reserved by the copyright owner. radiometricaspectw, e will discussthe fundamental ax e are modeled and characterized by a number of the relationshbipetweetnheimageintensitayndradiancferom anobject.Basedonthisrelationandimposedphysical lens distortion parameters. constrainttsh,emotion The image and object space coordinates of the points equations of image intensity will be are related by the collinearity condition in which the image derived for typical physical processes such as moving vector is aligned with the vector from the perspective Lamertian surface, emitting passive scalar transport, and center to the object space point transmitting passive scalar transport. These equations provide a rational way for reconstructing the geometric and kinematic properties of deformable bodies like fluids. -x,:+Sx: =2. M X:-X: (2.1) In general, the geometric, kinematic and radiometric approaches are closely coupled. - c X _- X ! where M = [m,j ] is the rotation matrix, )l is a scaling 2. Perspective Projection Transformation from 3D factor. Algebraic manipulation of Eq. (2.1) yields the Space to 2D Image well-known collinearity equations (with the distortion Image-based measurement techniques extract data from 2D images and then map them into the 3D object terms 8x s and ax e) relating the point in the 3D object spacc. There is aperspective relationship between the 3D space to the corresponding point on the image plane, coordinates in the object space and the corresponding 2D --1 m/( X - X_ ) X coordinates in the image plane [I, 6-8]. Here, we discuss x' -x_,+gx' =-c =-c-- several tbrmulations of the perspective projection m3r( X - X_ ) -X-3 transformation. Although these formulations are --2 ' (2.2) re,r( X - X_ ) X equivalent, one may be more convenient to use than others X2--X_+t_X: =--c =--C._-- i- for a specific problem. The fundamental geometric m._7( X-X_ ) X problem in image-based measurements is to determine the where the vectors m t = (m_ ,m_, ,mj¢ )r and object space coordinates X = ( X i, X "-.X 3)r given the m e = (m:_.me:,me. _)r are the directional cosine vectors corresponding image (retinal) coordinates x = (x_,x-_)7-. along the x_-axis, xe-axis in the image plane, respectively. Figure I illustrates the camera imaging process. The lens The vector m_ =(m,j.m,:,m,, )r is normal to the image of the camera is modeled by a single point known as the plane, directing from the principal point to the optical perspective center (or tile optical center), the location of center on the optical axis. As shown in Fig. 1, the unit which in the object space is Xc=(X(,X:,X_) r. orthogonal vectors mt, m2,and m_ constitute an object Likewise, the orientation of the camera is characterized by space coordinate frame at the optical center X_ and three Euler orientation angles. The orientation angles and location of the perspective center are referred to as the _- = (-_, ,_-:,_-3 )r are the projections of the object space exterior orientation parameters. The object space point. position vector X - X_ in this frame. The elements of the perspective center and image point lie along a straight line rotation matrix m_ (i.j = 1.2. 3) are functions of the Euler for a "'perfect'" camera. This relationship is described by orientation angles tco,¢._ ), the collinearity equations, the fundamental equations of photogrammetry. On the other hand, the relationship m. =cosO coslc, m_: = sinm sinO costc + costa sinlc, between the perspective center and the image coordinate m_, = - cosm sin¢ cos i¢+sin cosin Ic, m:j = -cos 0 sin lc, system is defined by the camera interior orientation m2e = --sin09sin0 sin _ +cos 09cos/c, parameters, namely, the camera principal distance c and the photogrammetric principal-point location m2_= cos COsin 0sin IC+ sinCOcos _C, xt, = (x_,, y_ )r. The principal distance c, which equals ms_ = sin O, m.,, =-sincocosO, m_ = COS(1) COS_). the camera local length for a camera focused at infinity, is (2.3) the perpendicular distance from the perspective center to The orientation angles (co.0.t¢) are essentially the pitch, the image plane, whereas the photogrammetric principal- yaw, and roll angles of the camera in the established point is where a perpendicular line from the perspective coordinate system in the object space. The rotational center intersects the image plane. Due to the lens matrix M is an orthogonal matrix having the property of distortion, however, perturbations to the imaging process M -t =M r or m[Cms =b'_,. The scaling factor lead to departures from collinearity that can be represented by the shifts 8x _ and 6x: of the image point from its 2=-c/m3r(X - X ) is a ratio between the principal C "'ideal" position on the image plane. The shifts _x _ and distance and the projected component of the object space position vector X-X_ on the optical axis in -m_ directionW. henanobjectspacpeointX is on the focal where X h=(X;,X2,X¢,I)r is the homogenous coordinates in the object space, and Ph= A-tMh and plane mj7(X-X )=0, the scaling tactor becomes C infinite, i.e., 2 = _, which corresponds to the points at M h=(M -MX_ ) are 3x4 matrices. Although Eqs. (2.6) and (2.8) arc formally written as a linear relation infinity on the image (retinal) plane. between xh and X or X h, they are essentially non- The terms ax; and ax 2 in Eq. (2.2) are the image linear because not only the lens distortion is a non-linear coordinate shifts induced by the lens distortion. The lens function of x, but also the scaling factor distortion terms can be modeled by the sum of the radial distortion and decentering distortion [9- I0] 2=-c/mJ(X-X ) is not a constant in general. • C a.r'=ax[ + ax,5 and ax e=a.r_ + a.r_, (2.4) Nevertheless, because the lens distortion is usually small, where, assuming that the optical axis of the lens is its effect can be corrected by using an iterative scheme. perpendicular to the image plane, we have Hence, Eqs. (2.6) and (2.8) can be treated as a quasi-linear a.'hI = K;( ";_' --"r;I' )re + K,( x1,-x;,; )i.4 , system at each iteration. Without the lens distortion, the a.r =K,(,,-:'-.q,)r"+X:(.r:'-.q,J," collinearity equations describe the ideal perspective projection. Eq. (2.8) is particularly suitable for utilizing ax5 = e,[ ,.2+ 2(x;,_xl ' )2]+ 2¢( x;,_xl ' )(x2,_.r_, ), useful results of classical perspective geometry to construct projective geometric invariants. a,r_ = _ / r2+2( xe'-x_, )2]+ 2P,( x"-x_, )(x2'-.r_ ) Furthermore, Eq. (2.2) can be re-written as a form ,.: =( x"-xl, )"+( .r:'-x_, ):. (2.5) suitable to least-squares estimation for the object space Here, K; and K2are the radial distortion parameters, P; and coordinates X, P2 arc the decentering distortion parameters• and .r_' and W;; ( X-X_ )=0 (2.9) x-" arc the undistorted coordinates in image. When the W2r(X-X_ )=0 lens distortion is small, the unknown undistorted where W_ and We are defined as coordinates can be approximated by the known distorted W_ =( x_-.ri,+a.r _)m3 +cm_ coordinates, i.e., x;'= x; and x:'= x2. For large lens (2.10) distortion, an iterative procedure is employed to determine W, =( x2-x_,+ax 2)ra, +cm 2 the appropriate undistorted coordinates to improve the As shown in Fig. 2. the vector W_ is on the plane spanned accuracy of the estimate. The following iterative relations by the orthogonal unit vectors m_ and m,, while W2 is are used: (x;')° = x; and (x2, )o= x2, on a plane spanned by m e and m 3. Geometrically (x_,)_+;=x; +ax; l( x;' )k,( x2' )k ] and speaking, W/( X-X_ )=O and W2r(X-X_ )=O (x2, )k+; = x2+ax e[( x;, )k,( x2, )*] , where the describe two planes normal to Wt and We through the superscripted iteration index k is k--0, 1,2..-. optical center. Thus, Eq. (2.9) defines an intersection of The collincarity equations Eq. (2.2) can bc re-written these two planes, which is a line through the optical center in the homogenous coordinates in the image plane X_. For a given image point x =(x;,x: )r Eq. (2.9) is xh= (.rj,x: ,x: )r = (xI,.r2 ,1)r not sufficient to determine a point in the object space with Ax h =2MfX-X_) or Xh=2P(X--X:), (2.6) the three unknown coordinates X=(X;.Xe,X')r. where P=[pii]=A-IM and A=[a,j] isdcfined as Hence, extra equations associated with additional cameras and other geometrical constraints should be added for '1 0 -x;, seeking a unique least squares solution of X . In contrast A= 0 1 -x_+ax 2 (2.7) to Eq. (2.8), Eq. (2.9) does not include the scaling factor I ";+ ax; 0 0 -c /t. The collinearity equations Eq. (2.2) contain the The terse tensor form of Eq. (2.6) is camera parameters to be determined by geometric camera aiixil = 2m,/X j -Xf ), where the Einstein convention calibration. The parameter sets (oo,O,_',X/,X,e,X;!), for summation is used. The matrix-form and tensor-torm (_v,x, ;,xT_ ), and (K;, K,,_PI,P_)_ in Eq. (2.2) are the of the collinearity equations are sometimes convenient for exterior orientation, interior orientation, and lens distortion mathematical manipulation. Another alternative form of the collinearity equations in the homogenous coordinates is parameters of a camera, respectively. Geometric camera calibration is a key problem in quantitative image-based (2.8) Xh = /_Ph Xh' measurements and a specific topic in both photogrammetry andcomputevrision.Hereweonlybrieflyaddrestshis curve lies can be reconstructed. When two calibrated issueandreaderscanfind the technicadletailsof cameras are used. the 3D curve can be uniquely geometriccameracalibrationfromreferencesI.n this determined as an intersection of two different projective paperw, egenerallayssumtehatthecameraiscalibrated conical developable surfaces. Furthermore, a 3D surface andacompletseetoftheorientatiopnarametearnsdlens can be reconstructed as an envelope of a family of the distortion parameters of the camera projective developable conical surfaces obtained from (ogO,,I¢,X,(,X,:,Xi_,c,x_,,x_,K,,K2,P_,P: ) is known. images taken at different viewing angles. The motion field Analytical camera calibration techniques utilize the of the 3D curve can be obtained from a time sequence of the curve. collinearity equations and distortion terms to determine Generating Proiective Developable Conical Surface these camera parameters [6-8]. Since Eq. (2.2) is non- linear, iterative methods of least squares estimation have Consider a 3D simple curve C in the object space, and its proiection to the image plane and a plane P normal to been used as a standard technique for the solution of the the optical axis (parallel to the image plane), as shown Fig. collinearity equations in photogrammetry. However. 3. The coilinearity equations Eq. (2.6) are written as direct recovery of the interior orientation parameters could be problematic and unstable since the normal-equation- X- Xc = 2-J-fi Xh, (3.1) matrix of the least squares problem is nearly singular. The where -fi=p-t =[Pi/I=M-tA=MrA" When the singularity of the normal-equation-matrix mainly results from strong correlation between the exterior and interior camera parameters and the scaling factor are constant and orientation parameters. In order to reduce the correlation the lens distortion is fixed, differentiating Eq. (3.1) yields between these parameters and enhance the determinability dX = 2-IP._2 dr . (3.2) of (c,xp,yp). Fraser 19, 1I] suggested thc use of multiple where dX =( dX I,dX :.dX -_)r. dr =(dr z dr: )r .and camera stations, varying image scales, different camera roll angles and a well-distributed target field in three I-PH -PJ: dimensions. Nevertheless, the multiple-station, multiple- image method Ibr camera calibration is not easy to use in L: many engineering and scientific applications like wind tunnel testing where optical access tbr cameras is limited A constraint imposed on Eq. (3.2) is mJdX=O, and the positions of cameras are fixed. AbdeI-Aziz and Karara 1121 proposed a simple linear method for camera indicating that Eq. (3.2) actually describes the projection calibration, Direct Linear Translbrmation (DLT). Cp of the 3D curve C on the plane P orthogonal to the Scientists in computer vision and robotics have developed optical axis direction or m._. This constraint is equivalent various camera calibration schemes to achieve a fast to the constancy condition of the scaling factor calibration with an acceptable accuracy (a lower accuracy 2 =-c/m_T( X - X ) since the differential for a photogrammetric application). Tsai's two-step C method [13] is representative in computer vision, which dA = cm_rdX (cid:0)(cid:0)m J( X - X, )1: shows uses a radial alignment constraint to obtain a linear least m_rdX = 0 ¢:_d2 = 0. In fact, the constraint squares solution tor a subset of the calibration parameters, whereas the rest of the parameters including the radial 2=-c/m_rfX-X_ ): const, defines the plane distortion parameter are estimated by an iterative scheme. orthogonal to the optical axis direction or m3. As shown By circumventing the singularity problem. Liu et al. [141 in Fig. 3, the projected curve Cp on the plane P can be developed a robust optimization method for single-image, automatic camera calibration to determine the interior and reconstructed from the image and then the developable exterior orientation parameters and lens distortion conical surface D containing the 3D curve C can bc generated. parameters plus the pixel spacing ratio. The arc length element of the projected curvc Cp on the plane P is 3. Projective Developable Conical Surface Containing 3D Curve dSc_ = IdX I= 2-I I-fi._:t Ids, (3.3) In this section, we introduce the concept of proiective where t =dr/ds and ds =ldxl are the unit tangent vector developable conical surface and show how to reconstruct and arc length element of the image of the 3D curve C in this surface containing a 3D curve from asingle image. In the image plane, respectively. Thus, the unit tangent principle, a 3D curve in the object space cannot bc vector of the projected curve Cp on the plane P is completely recovered from a single image since information in one dimension is lost in the imaging dX P._:t Tc, - - -- (3.4) process. Nevertheless, using a calibrated camera, a " dS,-k IP_: t I projective conical developable surface on which a 3D tangent plane on the developable conical surface D is Note that the unit tangent vector To,' is independent of the given by scaling factor 2'. The curvature vector of the projected (X-X_)oNo(s)=O, (3.1 I) curve Cp on the plane P can be obtained by differentiating where ND(s)=Tcp×(Xc, ,-X_)/ITcP×(Xc e-X_)l Eq. (3.4) with respect to the arc length Sc+' is the unit normal vector to the tangent plane on the dT¢,, _ 2 -- d t-P3,_t I developable surface, which is independent of the scaling (p,:k-Tce --), (3.5) Key - dScv [P.¢2t 12 ds factor. Eq. (3.11) describes a single-parameter family of the tangent planes where the parameter is the arc length s where k =dt/ds = d:x/ds 2 is the curvature vector of the of the curve in the image plane. The projective conical curve image in the image plane. The curvature vector k developable surface, the envelope generated by the family can be expressed as k=_n, where Ic and n=k/Ikl of the tangent planes, is given by a system of Eq. (3.11) are the curvature and the unit normal vector of the curve and Eq. (3.12)115] image in the image plane, respectively. Furthermore, we ( X- X_ ).dNo( s)/ds =O. (3.12) prove Thus, the projective developable conical surface and d IPc: t l_ (P._: ) (3.6) associated geometric quantities such as the curvature, k )T(-P32 t ds IPs: t I tangent vector and normal vector in the 3D object space can be obtained by using measured image quantities given Hence, Eq. (3.5) becomes the camera parameters. 2' -- (P_: n )r( P__2t ) ] Reconstructing 3D cuta,e and Surface Kc,, =__I_P .¢2t I2 [p_:n-Tc '' t-P._:t I . (3.7) From a single image, we arc able to reconstruct the proiectivc conical developable surface containing the 3D The curvature of the projected curve Cp on the plane P is curve C rather than the 3D curve itself. Nevertheless, _cp =Kcv'Nc,," where Nee =Kcr/IKc, ' I is the when two calibrated cameras are used, as shown in Fig. 4, principal normal vector of the projected curve Cp. Thus. the 3D curve Ccan be uniquely determined by intersecting the ratio between the curvatures _'c,, on the plane P and the two projective developable conical surfaces associated with the different cameras. Interestingly, the developable _" on the image plane is conical surface intersection method for determining the 3D K'c,__=L"2 curve only requires knowing the correspondence of one r_ Ip._:tl 2 [P_2n-Tc" (-P_:n)r(P_:tt)]l'Nc''lPe-- distinguished point such as an end point of the curve. Furthermore, the developable conical surfaces can be (3.8) used to reconstruct a 3D surface in the object space. As Clearly, Eq. (3.8) indicates that ,vc,'/t¢ is proportional to shown in Fig. 5, the developable conical surface the scaling factor 2,. containing the contour of the 3D surface can be After the unit tangent vector To," is obtained from the constructed. Here the contour is a set of points on the 3D surface at which the surface normal is also the normal of image, the projected curve Cp on the plane P is readily the developable conical surface. When the camera is reconstructed by moved to a number of known positions through a rotational and translational transformation (rigid-body Xc_, = Xcpo + I ,S(p Top(Sop)dSc_ . (3.9) motion), a family of the developable conical surfaces can The initial position Xo, . on the projected curve Cp, in the be obtained. The 3D surface is generated as an envelope of the family of the conical surfaces. Instead of moving object space is often chosen at the end point of the curve. the camera, the 3D surface can be rotated around a fixed Eq. (3.1) gives Xce,, - X _= 2,-JP xho , where axis such that a family of the conical surfaces can be Xho =(X/,.X_,I) r is the homogenous coordinates of the obtained using a camera at a fixed position and viewing angle. From a computational viewpoint, this method may corresponding image point to Xcr .. Substituting Eqs. not be the most efficient since the intersection and (3.3) and (3.4) into Eq. (3.9) yields a ray vector directing envelope of the developable conical surfaces has to be from the optical center X_ to a point Xc, , on the determined. However, this method is to great extent immune from the ambiguous correspondence problem in projected curve Cp stereovision. Xc_, -X_ =2,-I(Pxho P¢,_tds). (3.10) Recovering Motion Field o[3D Curve -- .+. Jl'*_ -- 0 Alter two or more 3D curves in the object space at A family of the projective rays through the optical center successive instants are reconstructed, we can estimate the X c given by Eq. (3.10) generates a projective developable motion field U( X ) of the 3D curve that is defined as conical surface D that contains the 3D curve C. The dX the curvature K of the filament along the binormal U( X )=-- (3.13) dt direction vector B [17] The curve is given by X = X[S(t),t], where t is time U( X )o¢ _B. (3.19) and S( t ) is the arc length of the curve in the object space. Overall, the physical constraints tbr a specific application are necessary for recovering the correct motion field and Measurements give the temporal and spatial difference associated physical properties of the 3D curve. between two curves at two successive instants t¢ and t, (the time interval At = t: -t_ is small) 4. Perspective Projection under Surface Constraint As,X =XIS(t: ),t:l-X[S(t I ),t_]. (3.14) In general, mapping between a point in the 3D object Reconstruction of the motion tield of the 3D curve from space and the corresponding image point is not one-to-one. Nevertheless, as shown in Fig. 6, under a given surface AstX is a non-trivial problem since the point constraint, a point on the surface has the one-to-one correspondence between two sequential images is not correspondence to the image point. In this section, wc known without using distinct targets on the curve discuss the geometric relationship between the surface in especially for an elastic curve experiencing large and the object space and the image plane. This topic is closcly complicated deformation. related to some applications in experimental fluid The motion field of the curve is constrained by the mechanics and aerodynamics such as reconstruction of underlying physical mechanisms behind the motion and complex Ilow topology from images of surface oil deformation of the curve. In general, reconstructing the visualization and laser-sheet-induced fluorescence motion field is fornmlated as an optimization problem of visualization. Consider a surface in the object space given the functional by J[U( X )]=IIA._,X-U( X )AtlI---_ min (3.15) X _= F( X I, X2 ). (4.1) subject to relevant physical and geometric constraints When Eq. (4.1) is imposed on Eq. (2.9) as a surface G,[U(X)]=O, (i=1.2,...) (3.16) constraint, the perspective projection transformation Eq. and the suitable boundary conditions. Without the (2.9) isreduced to sufficient constraints, the solution to the optimization (wHw:_-- wl._wel )Xt +(w_:w:_ - wl._w,_,_)X " problem may not be unique. Also, the imposed physical constraints serve as a bridge connecting image-based = w,._W/r X, - n'l._W2zX, measurements with the physical quantities in a specific w/iX_ +wl:X -"+wI._F(XI.X: )= W/rX . (4.2) problem being studied. where wj (i = 1,2 and j = 1,2,3 ) arc the elements of the In the simplest case in which the curve is rigid, the rigid-body motion field is expresscd as vectors Wt =( w_.wj:,wj, )7 and W2=(w:j,w::,w:_)r. U( X )=U o+ _,x( X-X o), (3.17) For the given surface equation X _= F( X _,X'- ) and the where Uo and .(2o are the constant translation velocity known camera parameters, the coordinates (X _,X: )r and angular velocity, respectively, and Xo is the rotational can bc obtained from the image coordinates x =(x _,x: )7 center of the curve. Because Uo and .(2, together contain by numerically solving Eq. (4.2). Thus. the coordinates only six unknown constants, it is easier to solve the X=(X_,X-',X -')7- in the object space can be optimization problem. A slightly complicated case is that the curve is stretched in three fixed directions in addition symbolically expressed as a function of the image to the constant translation and rotation. In this case, three coordinates x = (xj,x-' )7, that is, stretching constants are added, and thus the total number X =fs(x). (4.3) of the unknowns in the optimization problem is nine. In fact, Eq. (4.3) is a parametric representation of the Next, we consider a highly deformable material line surface using the imagc coordinatcs x =(x_,x-' )r as thc convected in an incompressible and irrotational flow. In this case, the physical constraints arc the solenoidal and parameters. Generally, the function fs(x) cannot be irrotational conditions [16] written as a closed-lorm solution except in some special V.U( X )=O and VxU( X )=O. (3.18) cases such as a plane and a cylindrical surface. A vortex-filament in an incompressible and irrotational Differentiating Eq. (2.9), we have flow isan interesting example since the filament driven by dW/X + W/dX = dW/X_ not only mean flow, but also self induction is no longer dW 2rX + W27dX = dW 2rX, . (4.4) passive and the motion field is directly related to the geometric features of the filament. In this case, the When the lens distortion is fixed, dWtT= dx_ms r and induced motion velocity of the filament is proportional to dW2 r = dx2m_ r hold. Then, substitution of curve on the image plane given by a parametric form dX ._=(OF/OX I)dX 1+(OF/OX 2 )dX 2 into Eq. (4.4t x(t)=(xJ(t),x:(t)) T and the corresponding 3D curve on yields the surface X(t)=X(x(t))=fs(x(t)), where t is a parameter (e.g. time). The length of an arc bounded the dX '- points corresponding to the parametric values t= t,, and where t=t I is o=(Wll +w_._3F/OXI wI, +wI_OF/OX 2 _w,i + w_,._3F/OX I w,_:+ w2_3F/OX: (4.6) S= ["[g,_/3(dx"/dt)(dx_/dt)] _/2dt. (4.15) ql h Furthermore, the differential dX J can be expressed as a The angle of two 3D curves at the intersecting point on the function of the image coordinatcs dx = (dx I,dr"_)T surface can be calculated based on the image quantities. dX -_=(dF/dx I )dx I+(dF/dr'- )dx:, (4.7) Consider two image curves x(t)=(x_(t),x2(t)) r and where x( t) = (xJ+(t), x"+(t))r • The tangential vectors of the two dF OF OXJ OF OX: 3D curves on the surface are - + .(ct = 1,2 ) (4.8) dx a OXI Ox" OX : Ox'_ dX (.rJ(t),x: (t))/dt = OX/&r_dx" / dt and Combining Eqs. (4.5) and (4.?), we have dX(x_*(t),x-'*(t))/dt = OX/Ox_+d__+ / dt. Thus, the angle dX = m.,r( X_ - f s )Qdx (4.9) yof intersection is where g_/j(dr"/dt)(dr/_+/dt) cos 7= _]g_/_(dr_,/dt )(dx, /dt )_/g,tfldx,,+/dt )(dx /_./dt ) -Q= ( ddtF.i ' ddxF2)/Qm-_Ir(.X¢-fs) _. (4.10) (4.16) The area of a domain H on the surface can be expressed in Eq. (4.9) provides a fundamental relation between the the image coordinates differentials dX on the surface and dr on the image plane. The matrix Q is a lunction of the image A(H) = [[,_-ff dx Idx:, (4.17) coordinates, the camera parameters and the geometric U where U is the domain in the image (x_,x :) plane properties of the given surface. On the other hand, we notice corresponding to the domain H on the surface in the object dX =(OX/_x I )dx j+(OX/Ox 2)dx2" (4.11) space and g is the determinant g =1ga/J I. From Eqs. (4.9) and (4.11), we obtain the following Example I: Plane equality The plane constraint is a simple, but very useful case (OX/OxI.OX/Ox "-) =mjr(x_ -fs )Q" (4.12) in which the vector function fs(x ), the matrices Q and The element dS of the arc length of a curve on the surface can be explicitly expressed as a function of the known can be determined from Eqs. (4.11) and (4.12) from the camera parameters and the measured image coordinates. image coordinates. We know Many aerodynamic flow structures are observed on a plane dS: = IdX I= g,,l_dr"d_ "_, (4.13) or a near-planar surface. Planar laser sheet Iqow visualization is just a typical case of the plane constraint. where OX OX In addition, a polyhedron consists of a number of the g_ = Ox,------O7.x_ (c_,fl = 1,2 ) (4.14) planar faces. Consider a plane in the object space is the so-called metric tensor in classical differential X _=aiX I+a,x:+a_. (4.18) geometry [I81. The summation convention is used in Eqs. This plane is defined by the vector a =(a I,az,a_)r (4.13) and (4.14). The quadratic differential form Eq. related to the normal vector of the plane. In this case. the (4.13) is the first fundamental form of the surface in which matrix Q in Eq. (4.6) is the image coordinates are the parametric variables. In the case of the perspective projection transformation, gaff Q=I w''+w'-_a' w'-'+w'-_a" )-. (4.19) 1,1'21 + W2d a l w22 + W23 a2 may be properly named as the perspective metric tensor The function fs(X) in Eq. (4.3) has a closed-form that is a function of the image coordinates, the camera solution parameters, and the properties of the given surface. The first fundamental form Eq. (4.13) allows us to measure the basic geometric quantities on the surface in the 3D object space from the image quantities. Consider a where Q-s! / //Q,,,)j (4.20) = . (4.27) a T Q (w,2P,,c°sq_-ws,p,,sinq) ws3 1 (%:p,, cosec- I%p,, sin _o w2_ where Another differential is dp =0. Note that the expressions ( T of fs(X), Q and Q lbr a spherical surface can be also i= Ws Xc - Irr_a_ (4.21 ) analytically derived, but they are so tedious that we do not T W2 X c - w:j ax present them here. Now the matrix Q in Eq. (4.10)is 5. Perspective Projection of Motion Field Constrained on Surface = (4.22) O-' (aj. a_,)Q-S After discussing the geometric relationship between a surface in the object space and the image planc, we study Example 2."Cylindrical Surface kinematics under the surface constraint, that is, the A cylindrical surface isanother case where fs (x), Q perspective projection of a motion field on a surface. and Q can be cxplicitly expressed. For the sake of Consider a dynamical system convenience, a transformation from the Cartesian dX i=U( X ), (5.1) coordinate system to the cylindrical coordinate system is dt used, i.e.. where U( X ) =(Us.Ue,U +)r is a motion field in the 3D X=( X_,X".x_ )r=(pcos_o, psinq),z) T, (4.23) object space and t is time. A surface constraint imposed where p is the radial coordinate, (p is the polar angle, and on the motion field Eq. (5.1) is Z is the axial coordinate. The differential dX is X 3= F( XI.X 2). (5.2) Under this surface constraint, U( X ) should be parallel to (,'os_ - p sin _ Oy dp ] the surface, which obeys the orthogonality condition dX =lsin _ pcos_ 0 Ida1. (4.24) N,. .U( X )=O. (5.3) t0 0 S X<I: ) where N =(3F/OXS,OF/aXe,-I) r is the normal For a cylindrical surface constraint P CO#1SI., = PC = vector of the surface. Under the surface constraint Eq. solving Eq. (2.9) lot ¢,0 and z, we have fs(x) as a (5.2), Eq. (5. I)is effectively reduced to a 2D system function of the image coordinates and camera parameters f s(x )= (p,, cos _0,p, sin q),z)r , (4.25) where X 2 U, IX_ X: FfXS.X2)] b/,, +xlt, +b:- b;b; In fact, Eq. (5.4) describes an orthographic projection of COS _ =- I_f+b7 the motion field Eq. (5. I) onto the plane (X s,X : ). From Eq. (4.5), the dynamical system in the image plane, which +_4t,7t, +b/- t,,< is corresponding to Eq. (5.4). is sin _o= (5.5) "= w_(( wnp,, cos_+ %z p,, sin _-W, rX, ), u=--_t( x-' m.,r(X,-fs) U:lfs (x)] o bs = Po( wHwex - ll':s%._ ), We call u =dr/dt =d/dt(xS,x: )r the optic flow in the b, = p (%_.W_v- w_,_u,'/._). image plane. The oplic lqow, a term first used in computer vision, is defined as the velocity field in the image plane b_=u' vl_lz TX_-%+W, TX,. that translorms one image into the next image in a There are two solutions for fs(x ), which are sequence. If Eq. (4.2) gives a one-to-one topological corresponding to two intersecting points between a mapping (homeomorphism): fxs,x_)_ (X _,X : ), the perspective ray and the cylinder. For a non-transparent topological structure of the dynamical system Eq. (5.5) in solid surface, a camera only sees one intersecting point at the image plane is equivalent to that of Eq. (5.4) on the thc surface lacing the camera and hence fs(X ) is one-to- surface in the object space when Q has the lull rank of 2 one. The differentials in the cylindrical coordinate system and m3r(x< -fs ) is not zero. Figure 6 illustrates this are related to the image coordinate differentials by the point. The problem is to recover two components of the |ollowing relation motion field (U_. U2)r using Eq. (5.5) from the measured d( (0,z)r = m iT( X c - f s )Q-J dr , (4.26)

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.