| consists of
ides greater
ymogeneous
tes on three
» a set of at
g this linear
nong the 27
t equations.
y equations.
n 3D model
"onsider the
n matrix, P,
lute ground
ositions and
of the three
irs of image
ains why P4
1 elements in
jy a set of 27
tions (7) per
nts (Ts), and
ber of points
y three of the
independent.
used as input
Edward M. Mikhail
for one of the nonlinear solutions that contain constraints among the parameters. These refined solutions are discussed
in the remaining subsections.
2.2.2 Constrained Solution with Four Dependent Equations (Model 1). There are four constraints among the 27 T
coefficients in Equations (7) that are commonly used. The first is that the 27" (i.e., the last) element of the vector T is
set equal to one. The next three are that the determinants of three 3 by 3 matrices, constructed from the T's, are equal to
zero; see [Hartley, 1997].
Although this solution with three constraints tends to alleviate some of the instability problems, the solution still is not
rigorous because it uses four algebraicly dependent equations. Since it is not possible to linearize with respect to the
observations (the corresponding coefficient matrix becomes rank deficient), the objective function to be minimized is
the sum of the squared errors of each equation (7) instead of the sum of the squared residuals to the observations.
2.2.3 Constrained Solutions with Three Independent Equations (Models 2,3). There are two models to be
discussed that use three independent equations per point. Both of these two models use the first three of the four
trilinearity equations; i.e., Equations (7). Since the three condition equations are algebraicly independent, we linearize
them with respect to the observations in order to obtain a rigorous solution.
Both models carry 24 parameters with six constraints since there should be only 18 independent parameters. In Model
2, the parameters are 24 of the T coefficients (T; through 75,). In Model 3, however, the values for the elements of Pa
and P,3; (aj; and bj) are computed from the initial estimates of the 27 T coefficients, and are used as the actual
parameters in the adjustment.
The six constraint equations in terms of T coefficients are found in [Faugeras and Papadopoulo, 1997]. The six
constraint equations in terms of the elements of P, and P,; are described in [Hartley, 1996]; however, since the
reference does not give them explicitly, they are listed below:
G =a,a,+a,a,, T 4,0, 70 G, 2450, t 450, t d5d,, 20
G, =Jau + a, +ay, —1=0 (8)
G, 54-120
G, 2a4,, t 4505, t d4,, 20
G, 24b, + bi, + bi, —1=0
2.3 Collinearity Model (Model 4)
All published derivations to trilinearity use the two scale factors,
allow more than one value for 7, when in theory there should be
only one. The following derivation, called "relative collinearity"
uses a single scale factor, 7, and four algebraically independent
equations. As in the trilinearity derivation, we compute the 3D
projective object coordinates from image 1, and then project the
object point into images 2 and 3; see Figure 2. This leads to the
following four condition equations:
(A. VA, I)
Figure 2. Relative Collinearity Model
F, 2x (ax as yas, t a,t)- (a,xaj,y aj, * a,,1)20
Fz y (as, X a4, y - a, +ayut)— (ay X+a,y+ay, +a,t)=0
”
y =X (by x + by, y+ by, + byt) (bx +b, y +b, +b,1)=0
F, = y (bx t b, y + ba Tbat)- (bx t by + by, *tb4t)z20
(9)
These four condition equations are independent; however they contain the additional unknown parameter 7, which is a
unique scale factor for each point observed. The other parameters involved are the 24 total elements of P,» and P, (aj
and b;), which are computed from the initial estimates of the 27 T coefficients. Note that the T coefficients are
estimated from the linear solution of the trilinearity equations, (7).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 587