As can be imagined, the resolving of point correspondences for
EO takes quite some computation time if all point permutations
on the EO device are considered. A total of 120 (or 5!)
resections need to be performed for a 5-point array before the
correct solution is known. On a Pentium II/266 PC, the EO
device detection takes two seconds whereas the scanning of a
single high-quality image (1536 x 1024 resolution) is performed
in only one second. Though computation time is usually not
critical in off-line VM, an optimised detection algorithm has
been implemented in Australis to further enhance processing
speed. If the targets on the EO device are labelled in a specific
order, the correct EO can usually be achieved with only four
resections. In the adopted labelling scheme, the point closest to
the centre of gravity of all EO targets (the raised point in Figure
4) is initially labelled, and then the remaining points are
incrementally labelled in a clockwise fashion. EO devices
which are labelled as such are detected within one second.
4. POINT CORRESPONDENCE DETERMINATION
USING EPIPLOAR GEOMETRY
4.1 Epipolar Plane Angle Geometry
Epipolar geometry can be employed to locate homologous
image points in a multi-image VM network. Therefore,
reasonable approximations of the EO parameters for all camera
stations are required, this being provided via the EO device. In
the literature, various techniques having different characteristics
in relation to speed and robustness can be found for image point
correspondence determination (e.g. Furnee et al. 1997, Chen et
al. 1993) All techniques, however, must deal with
correspondence ambiguities that can easily arise in the case of
only two images or in multi-image networks where there is a
dense array of targets (real and false). The technique adopted
for Australis is based on epipolar plane angles and is similar to
that proposed by Sabel (1999). It is fast and capable of
accounting for lens distortion effects in correspondence
determination, which is often critical when using modern CCD
cameras.
The basic principle of the epipolar plane angle method is that
rays of corresponding image points subtend identical angles to
any epipolar plane, as illustrated in Figure 6. Epipolar planes
always contain the projection centres of both camera stations. In
order to calculate the required angles for image points of both
images, an arbitrary epipolar plane is first defined. The
normalised vectors n; and n, are then computed with the applied
constraints that both the vectors and the baseline are normal to
each other, and that n, is contained by the defined epipolar
plane. The corresponding epipolar plane angle for each image
point in both images can then be calculated:
y z-R y (3)
ee
A =arctan(ñ, +, h, 7) (4)
Here, x; and y; are the coordinates of an image point corrected
for lens distortion, R is the rotation matrix for the current
camera station and c is the corresponding focal length.
4.2 Resolving Ambiguities
An image point can have an infinite number of homologous
points in the second image, the applicable constraint being that
they all have the same epipolar plane angles, or in other words,
they are positioned on epipolar lines. When lens distortion is
taken into account, the epipolar line transforms to a curve. Such
correspondence ambiguities can be resolved or at least greatly
reduced by using a third image. Consider the case of Figure 7,
where the point I, in image 1 has two possible homologous
points in image 2 (J; and J,), because they are positioned on the
corresponding epipolar line in the image. The two object points
P, and P, can be computed and are projected into the third
image. If one of these projected points is coincident with an
existing image point (K,), within a given tolerance, the
ambiguity is resolved and P, is then the correct 3D object point.
In the resection-driveback image-point measurement process,
where the object point coordinates and EO are utilised to predict
image point location, lens distortion has to be considered for the
back projection to deliver correct image point locations, and
therefore correct correspondence determination. An inverse
function for lens distortion is therefore needed. Because no
closed-form of this function exists, a numerical inversion that
typically converges within a few iterations is performed.
43 Correspondence Determination Algorithm
The entire correspondence determination process within
Australis, which was designed for highly-convergent VM
networks where most images 'see' a large portion of the target
range, is briefly outlined below. The process takes as input
several images with known EO and their corresponding image
point measurements.
Figure 6. Epipolar plane angle geometry.
264 —
CC
er
va
an
ar
in
cli
le:
in:
an
on
as
se
un
co
Th
wit
the
par
sen
for
the
stra
the
det
auti
me
can
Cor
Sca
alg
that