HIGH
RALIA
photogrammetric
imagery, which
data suppliers do
| surface models
re the geometric
are presented of
lied to two very
a, showed height
t both geometric
at results cannot
] photography (a
t al, 2002a).
hich investigated
onos images. In
used to constrain
erent geometric
straint, based on
id can be applied
oned to epipolar
d on the affine
ich like rational
space coordinates
ained by limiting
nadir lines in the
| to two Ikonos
vantages of each
AODEL
nilar to the RPC
es to object space
sensor model or
n of the model
3D object space
int i is expressed
(1)
er image, these
uniform scaling
) projections, one
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
scaled-orthogonal and the other skew-parallel. With high
resolution satellite imaging systems such as Ikonos and
Quickbird with narrow fields of view, the assumption that the
projection is parallel rather than perspective has been shown in
practical tests to be sufficiently valid (Fraser et al., 2002b). In
the reported implementation of the affine projective model, all
model parameters are recovered simultaneously along with
triangulated ground point coordinates in a process analogous to
photogrammetric bundle adjustment.
3. DATA USED IN THIS STUDY
Two pairs of satellite images were used in this study. Firstly a
stereo pair of Ikonos images, resampled according to epipolar
geometry by the data supplier and covering a 7 x 7 km area
over the city of Melbourne, Australia, was selected. The terrain
does not vary significantly in the area, with the lowest point at
sea level and the highest point at about 50m above sea level.
The Central Business District is located at the centre of the
area, and contains buildings up to 250m tall. About 15% of the
imagery depicts water at sea level. One of the images from the
stereopair is shown in figure |.
Figure 1. Ikonos image of Melbourne
The second image pair used in this study was a non-epipolar
Ikonos stereopair of San Diego, USA. Although this image pair
covered a wide ground area, a much smaller area of
approximately 7km x Skm was extracted. This sub-sampled
region was chosen in order to provide a very different test area
to the Melbourne data. Consequently it features mountainous
terrain, vegetation land cover and no urbanisation. One image
of this stereopair is shown in figure 2.
Figure 2. Ikonos image of mountainous region near San Diego
For each data set, the parameters of the affine projective models
were calculated using GPS-surveyed ground control points.
Most of the ground control points observed were road
roundabout centres, easily measured in image space by taking
the centroids of ellipses fitted to multiple edge points around
each roundabout. The remaining control points were. road and
building corners and other distinct features conducive to high
precision measurement in both object space and image space.
4. GEOMETRICALLY CONSTRAINED IMAGE
MATCHING
4.1 Background
In any image matching process there are generally three key
steps: selection of candidate points; definition of search space;
and, comparison of similarity measures. The most important, in
terms of practical implementation, is the definition of the search
space. By choosing the appropriate search space, computation
time is kept to a minimum and the potential for finding blunders
is reduced. In typical image space matching the search space is
a two dimensional area centred on the pixel being matched. The
search area has to be large enough to ensure the correct match
can be found, but not too large that the processing time
becomes computationally absurd. Thus, any way of reducing
the search space, but retaining the guarantee of the existence of
a correct match, is a significant improvement to any matching
algorithm. This is what geometric constraints aim to achieve.
4.2 Epipolar constraint
The most common geometric constraint used in image matching
is the epipolar constraint, which allows the search space to be
reduced from a two dimensional area to a one dimensional line.
By reducing the search space in this way, the speed of matching
algorithms can be increased by an order of magnitude, and the
chances of finding blunders is greatly reduced. With aerial
photography, epipolar lines in stereo images are usually
determined from a knowledge of the EO parameters. With high
resolution satellite imaging the EO is generally unknown,
meaning that the imagery must be purchased in its normalized
(i.e. epipolar projected) form.
Matching points in a stereopair of images that are aligned to
epipolar geometry is a simple task since it only involves a one
dimensional search space. However, by taking advantage of the
epipolar geometry it is assumed that the alignment to epipolar
geometry is error-free. Since the normalization of the Ikonos
imagery has been carried out by the supplier of the data (Space
