* There will be merely small rotation angles among the
stereo images. The effects of the Martian rotation are
negligible.
* The camera position and pointing information has an
excellent relative accuracy. Accordingly the valuation of
approximate values is more reliable. The area of
candidate search will be reduced. The correct
estimation of a triple of conjugate points can be
valuated in the object space. The collection of
conjugate points of images sampled in different orbits
makes a previous bundle adjustment essential.
* The search region of stereo candidates can be limited
to small areas along the track.
Nevertheless the matching algorithm should also deal
with images of overlapping respectively intersecting
image sequences. In particular the data of the WAOSS
sensor will have great overlapping regions especially at
the Martian poles. Using overlapping images puts high
constraints on the orbits to choose and high requirements
to the bundle adjustment. In addition to use overlapping
images the combination of HRSC and WAOSS data and
the use of the HRSC photometry channels is conceivable.
The redundant information should enforce the accuracy
and the reliability especially at the poorly textured Martian
regions.
Consequently the matching procedure requires a multi
resolution and a multi image matching solution. The
chosen matching strategy comprise a feature based and
an area based matching approach.
Expected poor texture images with only less features lead
to the suggestion to use a fixed raster with chooseable
grid size for matching. This approach has especially two
advantages. First the relationship among neighbouring
rasterpoints is well defined. Thus blunder detection is
easy. In the second place regions with poor texture will be
covered and there will be enough matched points to
compute object points and to achieve a good
representation of the terrain. In order to match distinct
points we use the feature based matching technique too
(Fórstner W., 1987). For the feature based matching the
approximate values must not be as good as for the area
based technique. Thus we use it also to optimise the start
points for the area based matching.
The results of the matching process will be conjugate
points and not DTM points. As a result of that there is the
feasibility to compute always easily new DTM's with actual
pointing data of the bundle adjustment process.
3.2 Preprocessing
All matching algorithm are more or less sensitive against
scale and rotation differences. Consequently the image
resolution and the image rotation differences of the
reference image and the stereo partners must be
adjusted. In order to keep locally the highest resolution
the images will be subdivided and adjusted part by part.
The size of the subdivided image parts will be chosen
dynamically and guarantee a locally nearly constant
resolution. To determine the different ground resolutions
and to get a first coarse localisation of corresponding
image areas the spacecraft navigation data will be used.
The scale adjustment takes place by a resampling
procedure with a low pass filter. In order to avoid image
distortion binomial filter will be used with the property of
shift invariance. The kernel size is variable and has to be
adapted to the scale differences. A further preprocessing
938
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
step is the generation of image pyramids. The image
pyramid starts at the level with the highest common
ground pixel resolution. The decimation steps are chosen
to the power of two.
3.3 Approximate values
An important aspect is the availability of good and reliable
approximate values, to assure sufficient convergence of
the subpixel matching. Above all for the least square
matching method the correct pixel position of the discrete
pixel determination has to be better than two pixels. In
order to get these good and reliable approximate values a
hierarchical coarse to fine strategy by using image
pyramids is implemented. At the first pyramid level we get
the startpoints either from the conjugate points, used by
the bundle adjustment, or from the pointing data of the
cameras. For this case a grid of anchor points is defined
in the reference image and will be transferred in the object
space. For all the grid points we compute the intersection
point of their line of sight and the Martian ellipsoid. Next
we search in the stereo image the scanning line which
defines together with the line of sight of the centre pixel a
plane which include the intersection point or is located
close to it. The corresponding pixel will be found across
the image line.
The approximate values will be used to compute the
coefficients of a polynom function to determine corres-
ponding feature of the candidate list of the feature
extraction and to detect incorrect matches.
3.4 Subpixel matching
In order to get subpixel accuracy a multiple image least
square matching technique is used (Tsingas V., 1991).
Analogous to the stereo least square matching a
minimising of the grey value differences of all image
patches is carried out. The estimation is performed by
using the affin transformation. If we have n image patches
there will be n*(n-1)/2 possible transformation
combination. The transformation parameters of all
combination are highly correlated. Therefore one of the
patches will be the reference patch. All transformation
between two patches can be described by the roundabout
way of the transformation to the reference image. With
that the number of transformations is reduced to n-1 with
6*(n-1) independent transformation coefficients.
4. SOFTWARE TESTS WITH PLANETARY IMAGE
DATA
Up to now, no image data which such a complex image
geometry is available. Therefore we test the software with
images of diverse planetary missions. In particular the
image data from the Clementine misson to the Moon and
the Galileo mission to the asteroid Ida are suitable to test
the matching software.
The Clementine data serves the main aspects for the
software tests:
* multiple images.
forward, nadir and backward viewing images.
poorly textured regions and surface discontinuties.
position and pointing data stored at SPICE-kernels.
different ground pixel resolution.
To perform the tests only a few software modules had to
be adapted. The tests yielded reliable results with high
"am ta^ NS
th "S fr^ fn
— PP d -— 9 (AS pA cp CC 22 ad I
fth
Fa u PM PN rm a —
IE AN 0 qu ey
