Escolar Documentos
Profissional Documentos
Cultura Documentos
Satoru Takahashiy and Bijoy K. Ghoshz Department of Systems Science and Mathematics, Washington University St. Louis, Missouri 63130, USA
y
saru@zach.wustl.edu,
ghosh@zach.wustl.edu
systems, canonical forms, Kronecker index, parameter identi cation, extended Kalman lter
Perspective Parameter Identi cation: Assume that we observe Z (t) in an unspeci ed interval 0 T ). The problem is to determine the extent to which the parameters AT C are identi able from the observed data.
Perspective problems have already been considered in the literature Kanatani 5] where the goal is to identify parameters of motion using a `charged coupled device' CCD camera. Generically it is known 2] that parameters can be identi ed up to orbits of a perspective group. The scalings on the parameters (AT C ) as a result of the group action is described as follows,
1 Introduction
An important and somewhat di cult problem in machine vision is to identify parameters of motion dynamics from observing projection of feature points on the image plane observed over time. The di culty arises from the fact that the observation function is nonlinear (perspective) and the underlying sensor (a CCD camera) is noisy. In fact, not all the parameters are identi able and in 2], the identi able parameters are characterized via orbits of a \perspective group". In order to identify orbits, we need to de ne a \canonical form" , 6] { 13], which is the main subject matter of this paper. To illustrate the main concept we consider a linear nite dimensional dynamical system of the form X_ = AT X Y = CX X (0) = X0 (1.1) where X is the state variable of dimension n and where Y is the output variable of dimension m. We assume furthermore that the output is observed only up to an unknown scale factor. Thus the observation function is given by Z : IRm ; f0g ;! IRIPm;1 (1.2) Y 7;! Y ] where we assume that m > 1. Note that the observation function is de ned for all nonzero vectors in IRm .
X : GL(n) P ;! P (P AT C ) 7;! (P AT P ; C P ; )
1 1 1
X : IR P ( AT C ) X : IR P ( AT C )
2 3 +
The perspective group is the direct sum of each of the three groups in (1.3), (1.4) and (1.5). As a result of the action of the perspective group on the parameter space, the space is split up into orbits. It may be trivially veri ed that parameters in the same orbit of the perspective group cannot be observed using (1.2). Furthermore it has been shown by Ghosh et.al. 1], 2] that generically the orbits of the perspective group are indeed observable, i.e. produces distinct outputs for at least some interval of time. In order to actually carry out the problem of observing the orbits, it is essential to parameterize the orbits by describing a chart with an appropriate set of coordinates and construct an observer that computes an estimate of the coordinates. In this paper we consider primarily the parameterization problem but also illustrate our results with observers that are derived from EKF.
as follows
0 1 0 1 0 d @ X A=@ b A+@ a b a dt Y
1 2
11 21
0X 0 f f f 0 0 0 1 B XY B B Y @ 0 f 0 f f 0 A B XZ + B 0 0 f 0 f f B YZ @
2 1 2 1 3 2 3 2 2 1 3
b3
a31
10 X 1 A@ Y A 1 C C C: C C C A
Z
2y 3 01 4 y 5=@ 0
1 2
y3
0 0 1 0 0 0 1
12 X 3 0 6 Y 7 0 A6 Z 7: 4 5 0
1 1
W1
(2.6)
If only the rst component of the observation function (2.1) is available, we can write
y1 = y3
1 0 0 0 0 0 1 0
2X 3 6 Y 7: 6Z 7 4 5
1 1
W1
(2.7)
(2.1) Let us denote A = (aij ) b = (b1 b2 b3 )T f T = (;f1 ; f2 ; f3 ) where i j = 1 2 3. The dynamical system (2.1) is a Riccati dynamics in IR3 , which has been illustrated in 3]. It is a class of quadratic motion models more general than a rigid ow which preserves shape, i.e., the shape of a planar surface remain planar, although the distance between two points on the plane may not remain constant. An important question, that is of interest in machine vision is to ask the following. and assume that the state vector (X Y Z ) is observed by the function x = X=Z y = Y=Z Z 6= 0 (2.2) which is typically the case under perspective projection, or x=X y=Y (2.3) which occurs under orthographic projection, to what extent are the motion parameters and the initial conditions X (0) Y (0) Z (0) identi able from the observation vector (x(t) y(t)) in a given time interval 0 T ) T > 0? In order to show the connection between the dynamical system (2.1), (2.2) with that of (1.1), (1.2) note that one can homogenize the state vector (X Y Z ) as X = X1 =W1 Y = Y1 =W1 Z = Z1 =W1 (2.4) and write (2.1) as
Z2
Note that the state vector in (2.5) and the observation vectors in (2.6) and (2.7) are elements of a projective space and the pairs (2.5), (2.6) and (2.5), (2.7) are of the form (1.1), (1.2). Likewise if we consider orthographic projection (2.3) as the observation function, then (2.6) is changed to the following equation
2y 3 01 4 y 5=@ 0
1 2
y3
0 0 0 1 0 0 0 0 0 1
12 X 3 6Y 7 A6 Z 7: 4 5
1 1
W1
(2.8)
In this paper we distinguish between two important cases, one in which the camera calibration is unknown and the other in which the camera calibration is known. If the camera calibration is known, one can in turn assume that the matrix C in (1.1) is known and can be chosen to take the special structure described in (2.6), (2.7) and (2.8). Otherwise the matrix C is assumed to be unknown.
2X 3 2X d 6 Y 7= A b 6 Y 6 7 fT 0 6 Z 4 dt 4 Z 5
1 1
1 1
W1
W1
3 7: 7 5
(2.5)
The observation vector (2.2) can also be written, by de ning x = y1 =y3 y = y2 =y3
(3.1) 0 0 1 1 2 3 4 where the parameters i , i , i = 1 2 3 4 and are arbitrary. Writing B as (b1 b2 ) we assume that 0
1
0 B B @
1 0
2
0 0
3
0 1
4
1 C C A
00 B1 B0 @
1 C C A
for some scalar constants 1 2 3 4 , 1 2 that can be determined from the pair A, B. If we consider a basis given by
Ab = Ab =
2 2 1 2
1 1
00 B AT = B 1 @0
0 B B @
0 0
1 2
1 0
3 4
0 0
1 2
1 C C A
00 B1 B0 @
0 0 0 0 1
1 C C A
(3.2)
where we assume that the pair in (3.2) is the representation of A, B with respect to the basis given by
(3.4) and the perspective system is given by x1 = 1 x3 + 2 x4 _ x2 = x1 + 3 x3 _ x3 = x2 + 4 x3 _ x4 = 1 x3 ; 4 x4 _ z = x3 =x4 : There are a total of 5 parameters for the Kronecker index pair (3 1). Finally if we now choose n = 4 and m = 3 and assume that the Kronecker indices are 1 = 2, 2 = 1, 3 = 1, it is possible to show that a suitable canonical form is given by 00 1 0 1 1 1 1 B 1 2 2 2 C C=@ 0 1 0 0 A 0 1 1 0 AT = B 0 3 3 3 C @ A 0 2 3 1 0 4 4 4 wherein, like before, the parameters i i i , i = 1 2 3 4 and j j = 1 2 3 are arbitrary. Denoting the columns of B := C T by (b1 b2 b3 ), the Kronecker structure implies that the vectors b1 b2 b3 and Ab1 are independent and A2 b1 = 1 b1 + 2 b2 + 3 b3 + 4 Ab1 Ab2 = 1 b1 + 2 b2 + 3 b3 Ab3 = 1 b1 + 2 b2 + 3 b3 for some scalar constants 1 2 3 4 1 2 3 1 2 3 that can be determined from the pair A, B. If we consider a basis given by
0 0 1 0 0
1 3 4 1
0 0
1 C C= C A
0 0 1 0 0 0 0 1
fAb ; b ; b b Ab ; b ; b b g:
1 3 1 4 2 1 2 3 1 4 2 2
If we now consider the group action (1.4) on the pair (3.2), we can assume that the trace of the matrix A is 0. Thus we can replace 4 by ; 3 . Therefore a canonical form for the perspective system (1:1), (1:2) is given by
00 B AT = B 1 @0
0
1 3 1 3
0 2 0 4 0 2 1 ;3
1 C C= C A
0 1 0 0 0 0 0 1 (3.3)
that is,
x_2 = x1 + 3 x2 + 4 x4 x_4 = 3 x2 + x3 ; 3 x4
Therefore, for the Kronecker index pair (2 2) we have a total of 7 parameters. If we choose n = 4 and m = 2 and assume that the Kronecker indices are 1 = 3 and 2 = 1, a suitable canonical form is given by
00 B AT = B 1 @0
0 0 1 0 0
1 2 3 4
1 2 3 4
1 C C= C A
0 0 1 0 0 0 1
00 AT = B 1 @0
0
1 4 1 1
2 2 2
3 3
where the parameters i , i , i = 1 2 3 4 and are arbitrary. Denoting the columns of B := C T by (b1 b2 ), the Kronecker structure implies that the vectors b1 , b2 , Ab1 and A2 b1 are independent and
Ab =
3 1
Ab2
b + 2 b2 + 3 Ab1 + 4 A2 b1 = 1 b1 + 2 b2
1 1
; 4;
1 C C= A
0 0 0
1 0 0
0 1 0
0 0 1
Note that for the Kronecker index (2 1 1) there are a total of 9 parameters. From the above remark one obtains the following result.
era calibration) : Let us consider the perspective dynamical system (1.1), (1.2) for the cases n = 4 m = 2, n = 4 m = 2 and n = 4 m = 3, for the Kronecker indices (2, 2), (3, 1) and (2, 1, 1), respectively. A canonical form for each of the three cases is given by (3.3), (3.4) and (3.5), respectively.
where fwk g 2 IRn is assumed to be a Gaussian white noise with mean zero and covariance R, i.e. Efwk wlT g = R kl , kl indicates Kronecker's delta and H = On 7 n 3n] 2 IRn (3n+7) . Here n 3n is a suitable matrix and its detailed structure is omitted. The EKF is described as follows. ^k+1jk+1 = F ( ^kjk ) + Kk+1 Zk+1 ; HF ( ^kjk )]
Kk
Fk
+1
Qk+1jk
Remark 3.2 : Note that for each of the above three cases, the group action (1.5) can be used to scale the matrix C to a suitable canonical structure, viz. a matrix with unit norm. The details about this have been omitted. 3.2 Parameter Identi cation using an EKF
Parameter estimation problems have already been studied using an EKF by many researchers in the past, see for example Kano 4] for motion and shape estimation under perspective projection. In this section we only discuss the case of estimating the seven parameters in (3.3). Discretizing the dynamical system (3.3) we obtain the following discretized dynamical system Xk+1 = Ad Xk zk = x2k =x4k (3.6) 01 1 T 0 T B T 1 +1 3T 0 2T C 4 C Ad = B 0 @ A 1 1T 2T 0 T T 1 ; 3T 3 where Xk = (x1k x2k x3k x4k )T and where T is the sampling time. Note that the parameters to be estimated are 1 2 3 4 1 2 3 . The three coordinates of a point observed by the camera are given by xk = x1k =x4k yk = x2k =x4k zk = x3k =x4k : (3.7) We now assume that, at each time instant k, a set of n points are observed with coordinates (xik yik zik ) for i=1 n. The coordinates are de ned as in (3.7) for each of the n points. We consider the state vector k to be given by k=( 1 2 3 4 1 2 3 x1k y1k z1k : : : xnk ynk znk )T 2 IR3n+7 and the associated state equation is given by (3.8) k+1 = F ( k ): In fact, it follows from (3.6) that we need only the observation values yik . Let yik be observed with an additive noise, and let Zk 2 IRn be the observation vector. Then Zk = H k + wk (3.9)
(3.10) Likewise an EKF for the dynamical system (3.4) and (3.5) may be described similar to that given by (3.10). Simulation studies on the EKFs are performed for the following parameters. For the Kronecker indices (2 2) parameters in (3.3) is taken to be Case (2 2) : ( 1 2 3 4 1 2 3 ) = (;0:2 0:1 1 0:5 1 0:3 1:5): (3.11) Moreover we set the sampling period T = 0:01, the number of points is chosen as n = 3 and (x10 y10 z10 ) = (0:05 0:3 ; 0:1) (x20 y20 z20 ) = (0:3 0:1 ; 0:06) (x30 y30 z30 ) = (;0:2 0:1 ; 0:1): Let the noise variance be chosen to be R = 2 I3 and R = 2 I6 with = 0:0001, respectively. The estimates of the parameters in (3.3) have been estimated and plotted in Fig. 1. The dotted lines denote the true values in (3.11) and the solid lines represent the estimated values with the EKF. It may be noted from the gure that the initial condition for the parameters chosen during the simulation of the EKF is chosen to be close to the true value. Otherwise for various other choices of the initial conditions, the parameters do not converge to the true value but either maintains a xed bias or diverges.
@T
= ^kjk
2X 3 d6 Y 7 6 7 dt 4 Z 5
1 1
W1
2X 3 6Y 7 AT 6 Z 7 4 5
1 1
W1
(4.1)
0.15
0.25
0.2
0.2
0.25
0.15
0.3
0.1
spective dynamical system (4.1) with the observation function (2.6) and assume that b 6= 0. Under the perspective group action (1.3), (1.4) and (1.5) using P of (4.2), a canonical form in the case of known camera calibration is given by the subset S de ned as follows:
0 200 400 600 sampling 800 1000
r1
r2
0.35
200
800
1000
0.05
S=
0.6 0.58 0.56 0.54 0.52 0.5 0 200 400 600 sampling 800 1000 0.48 0 200 400 600 sampling 800 1000
A b fT d
1.25 1.2 1.15 delta1 delta2 1.1 1.05 1 0.95 0.9 0 200 400 600 sampling 800 1000
0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0 200 400 600 sampling 800 1000
1.6
1.55 delta3
1.5
1.45
1.4
200
800
1000
Figure 1: Kronecker index (2,2) case with the observation function being (2.6) or (2.8) where we de ne b AT = fA d : T
4 A =A+b 4 b = (b1 =b3 b2=b3 1)T b3 f T ; A ; ( b ; d) ;( b ; d) where denotes the vector (p41 p42 p43 ). Here we can choose p4i = ;b T Ai =kb k2 , i = 1 2 3 where Ai is the ith column vector of A. It follows that the vector is unique and b T A = 0. Furthermore in order to get traceA ; ( b ; d) to be 0 we apply the group in (1.4) with = (;a11 ; a22 ; a33 + ( b ; d))=4 where aii , i = 1 2 3 denotes a diagonal element of A . The cases b1 6= 0 or b2 6= 0 can be described likewise. (Q.E.D.)
Proof of Theorem 4.1 : From the assumption b 6= 0 we assume that b3 6= 0 without any loss of generality. Then we may choose p11 = b3 and p44 = 1. The scaling on the matrix AT described by the group action (1.3) is given by
A b fT d
AT =
0 a B B ;b a a ; b a @
11 21 1 11
;f
2 21
;b a ; b a ;f
1 12 2
a12 a22
2 22
;b a ; b a ;f
1 13 3
a13 a23
b1 b2
2 23
;a ; a + b a + b a
11 22 1 13
1 C C A
(4.4)
2 23
p41 p42 p43 p44 In general the action of P changes the structure of the associated matrix C , but P of (4.2) preserves the structure. We now have the following result.
11
We set the sampling period T = 0:001, and the number of points is taken as n = 3 and (x10 y10 z10 ) = (0:2 0:1 ; 0:2) (x20 y20 z20 ) = (0:3 0:2 0:1) (x30 y30 z30 ) = (;0:1 0:3 ; 0:1): Let the noise variance be chosen to be R = 2 I6 with = 0:0001, respectively. Now using the EKF denoted by (3.10), parameters in (4.4) have been estimated and plotted in Fig. 2. Let the dotted lines denote the true values and let the solid lines indicate the estimated values. The simulation of the EKF is shown to be close to the true value. It follows from Fig. 2 that parameters are estimated with quite an accuracy.
a12 b1 a22 b2 A = 0 and 21 ;f1 ;f2 d ci = 1 (i = 1 or 2 or 3)g where ci is the ith element of c.
11
S=
A b fT d
traceA + d = 0
0a cT @ a
5 Conclusion
In this paper, we have introduced canonical forms for perspective dynamical systems under the action of a perspective group assuming both unknown camera calibration case and known camera calibration case. In the former case we show that the problem is related to the classical Kronecker indices. We also observe in this case that because the calibration parameters are not known apriori the associated perspective group describes orbits of \large" dimension. Hence the number of identi able parameters are quite small (7, 5 and 9 in our examples) compared to the total number of parameters. In the latter case, when the calibration parameters are known, 11 of the possible 16 parameters under perspective projection are identi able and has been identi ed in our example. We parameterize orbits of the perspective dynamical systems, and show via simulation that the parameters can be identi ed using an EKF. In this way, we provide a geometric framework to study the problem of parameter identi cation of a linear dynamic system with perspective and orthographic observations.
References
1] B. K. Ghosh, M. Jankovic and Y. T. Wu, Perspective problems in system theory and its application to machine vision, Journal of Mathematical Systems, Estimation and Control, vol. 4, no. 1, pp. 3 { 38, 1994. 2] B. K. Ghosh and E. P. Loucks, A perspective theory for motion and shape estimation in machine vision, SIAM Journal on Control and Optimization, vol. 33, No. 5, pp. 1530 { 1559, 1995. 3] B. K. Ghosh, H. Ianaba and S. Takahashi, Identication of Riccati dynamics under perspective and
4]
0 0.05 0.1 a11 a12 0.15 0.2 0.25 0.15 0.1 0.05 0 0.05 0.1
5]
0 2000 4000 sampling 6000 8000
2000
4000 sampling
6000
8000
6]
0.5 0.45 0.4 a13 a21 0.35 0.3 0.25 0.2 0.15 0 2000 4000 sampling 6000 8000
7]
0 2000 4000 sampling 6000 8000
8]
0 2000 4000 sampling 6000 8000
2000
4000 sampling
6000
8000
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0 2000 4000 sampling 6000 8000 0.55 0 2000 4000 sampling 6000 8000
0.05 0 0.05 f1 f2 0.1 0.15 0.2 0.25 0.3 0 2000 4000 sampling 6000 8000
0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0 2000 4000 sampling 6000 8000
0.2 0.25 0.3 f3 0.35 0.4 0.45 0.5 0.55 0 2000 4000 sampling 6000 8000
orthographic observations, IEEE Transactions on Automatic Control, September 2000, to appear. H. Kano, B. K. Ghosh and H. Kanai, Recursive Estimation of Motion and Shape Parameters of A Moving Rigid Body from Noisy Image Data under Perspective Projection, Proceedings of the Mathematical Theory of Networks and Systems, pp. 1039 { 1042, 1999. K. Kanatani, Group-Theoretical Methods in Image Understanding, Spring-Verlag, 1990. M. Hazewinkel, Moduli and canonical forms for linear dynamical systems III : the algebraic-geometric case, Geometric Control Theory (C. Martin, R. Hermann, eds.), Math. Sci. Press, Brookline, 1977. M. Hazewinkel and R. E. Kalman, On invariants, canonical forms and moduli for linear constant nite-dimensional dynamical systems, Lecture Notes in Econo. ; Math. System Theory, Springer-Verlag, Berlin, vol. 131, pp. 48{60, 1976. R. Hermann and C. F. Martin, Application of algebraic geometry to system theory : the McMillan degree and Kronecker indices of transfer functions as topological and holomorphic invariants, SIAM J. Control, vol. 16, pp. 743 { 755, 1978. R. E. Kalman, M. Arbib and P. Falb, Topics in mathematical system theory, McGraw Hill, New York, 1965. R. E. Kalman, Kronecker invariants and feedback, Ordinary Di erential Equations (L. Weiss, ed.), Academic Press, New York, pp. 459 { 471, 1972. R. E. Kalman, On partial realizations, transfer functions and canonical forms, Acta Polytechnica Scandinavica, Helsinki, MA31, pp. 9 { 32, 1979. V. M. Popov, Invariant description of linear timeinvariant controllable systems, SIAM J. Control, vol. 10, pp. 252 { 264, 1972. T. Kailath, Linear sytems, Prentice-Hall, Inc., Englewood Cli s, 1980.