The relationship between image coordinates on the left
(XL,yL) and (xR,yR) on the right is given by assuming an
affine transformation exists between the windows
(Foerstner, 1982). Furthermore, no information of the
object is taken into account in the matching process.
Matching is solely based on the intensity values of the
pixels and the assumed affine transformation. However, if
the collinearity conditions were to be used, the
transformation could take into consideration the shape of
the object. If some information about the object's surface
is availabale, then a sensible surface model can be
introduced and the model could be improved. This would
then constrained the matching to the surface model. In
doing so, image coordinates (XR,yR) are expressed in
terms of (xL,yL) using coordinates (X,Y,Z) on the object
surface.
Consider a window of size n x n pixels (where n is odd)
on the left image with its centre (i.e. the central point)
having the coordinates of (x ,y| ). It should be noted that,
‘central point‘ does not have to be the centre of the
window. The term is chosen merely for the convenience
of explanation. By using an appropriate surface model,
the corresponding position on the right image (xp,YyR),
where xR,yR are not neccessarily integers, can be in
terms of the central point on the left (xL,yL) and the
corresponding coordinates (X,Y,Z) on the surface. In
addition to the central point, the relationships of
neighbouring points, say, (xi -AxL , yL+AyL) on the left
image and (XxR+AXR, YR+AYR) are also needed. Values
AX| and Ay|_ are known while AxR and AyR are not known
but are defined by the relevant transformation between
the windows.
Use of the collinearity equations, would introduce three
additional parameters, (X,Y,Z) for the central point and
also for each of the neighbouring points. Supposing that
the six relative orientation parameters are known, then a
relationship can be established that relates (xL,yL) to
(X, Y,Z). By adopting a suitable surface model coordinates
of neighbouring points on the surface can be related to
the central point and the replacement of the affine
transformation is achieved.
2.1 Functional Model
The coordinates (xL,,yL,) of the central point, O, of the
window on the left image can be represented by :-
XL 7 f (Xo. Yo,Zo) ; yL 7 fy, (Xo.Yo,Zo) ... (ii)
where, xj and yj, are known; (Xo,Yo,Zo) are not known
(but needed) and an initial estimation can be obtained;
fx and fy, are the collinearity conditions. For a
neighbouring point, P, with shifts (Ax|,Ay|) from the
origin on the left image, then its coordinates can be
represented by :-
XL*AxL — fx, (Xp, Yp,Zp) N
YL+AYL =fy, (XP,YP,ZP) … (iii)
where Xp,Yp,Zp are the coordinates of the neighbouring
points on the surface. If the corresponding shifts on the
object are AX, AY and AZ, then eqn (iii) can be written
as :-
556
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
XL+AXL m fq (Xo* AX, YoAY Zo AZ)
YL+AYL = fyı (Xo+AX, YotAY,Zo+AZ) s (iV)
Supposing that matching is to be done for a flat surface
then a surface model can now be introduced across the
window to represent the surface. This is given by
(considering only the first order terms, i.e., a planar
surface model) :-
AZ = 2 None ny NAV)
ox oY
where (0Z/0X) and (0Z/dY) are the gradients of Z in the X
and Y directions respectively. These gradients define the
model surface and they are to be evaluated in the
solution. In turn, the terms AX and AY can now be
expressed in terms of AxL, AyL :-
X
AX 2 — Ax +—Ay … (vi)a
Sx D yo
oY oY
AY = —A —A … (vi)b
Ox. Carr YL (vi)
where (oX/ox;), (0X/dyL), (aY/oxL), (dY/dyL ) are derived
from the collinearity equations and Ax; & Ay] are
known shifts on the left image. By substituting eqns (vi)a
and (vi)b into eqn(v), AZ can be expressed as a function
of the shifts Ax| , Ay| :-
= 9Z X AX kdX ay 4 92 on ony
oX|ox Lay 'L av|ax Loy Lb
"I
This implies that AX, AY and AZ can be written in terms
of the known shifts Ax|_ and Ay, as well as the surface
gradient (0Z/0X) and (0Z/dY). Since AX, AY,AZ, (0Z/0X)
and (0Z/dY) represent the planar surface model of the
object, these terms are therefore common to both the left
and right windows. In other words, given a neighbouring
point on the left image whose are (x +Ax;, yL*AyL), its
corresponding coordinates on the right (xg+Axg, yR+A
YR) can be estimated via the planar surface model. Thus
eqn (i) can now be expanded to represent neighbouring
points by :-
ILXL+AXL,YL+AY| ) + N(X+AXp,y+Ay| ) =
IR(XR+AXR,YR+AYR ) ... (Viii)
Linearising eqn(viii) in xa, yp, Axq and AyR will yield :-
IL(X_+AXL,YL+AYL) + N(X+AX ‚y+Ayı ) =
IR° xR +axR%yR%+AYR®)
dlp dlp dig dlp
— Id —— d ——- dA — dA
ES XR «| 2: YR «2n XR + Wn YR
... (ix)
where
dxg, C
priori \
from a
manus
compu
(oZ/oX
olr/ox
x and:
Lookin
dxg a
solutio
matchi
dAyR
expres
AXR =
Substi
replaci
will giv
AXp =
Equati
AXL /
colline:
of xn
Hasc
only G
dAXg =
The p
obtaine
0AXR :
JG
0AXR €
oH -—
Similar
Ayg =