\GERY
1nover,
dlr.de
! Berlin,
orbit. The
y f Mars in
Hannover,
btained tie
te Sensing
ese results
[.
results of
of the tie
ent (Ebner
HRSC is
rated in a
n verified
1s of star
ition have
> HRSC is
] for the
sented. In
of the tie
ue shown
and some
2gy which
1gh image
observed
e interior
ible to use
| accuracy
eumann et
r operator
sise in all
icient as
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
similarity measure. Each image is divided into subareas to
ensure an cven distribution of the tie points over the whole area.
To reduce ambiguities and computing time the matching
location and a search space for the corresponding feature has to
be determined. The principle of the transformation from object
to linescanner coordinates is described in Dórstel, Ohlhof
(1996). Since no epipolar geometry exists for linescanner
imagery a feature in one image is transferred to the next image
via the collinearity equations for 3-line imagery (1) according
to the extended functional model of Ebner et al. (1994).
X—-X X X Ü + AX,
y- y, |=AM'(Ap,Aœ,Ax)D'(p,0,x)|| Y Y + AY, (1)
= 2} 12, +45
with
AX 0 Ax
AY, |z D(9,0,Kk)| Ay (2)
AZ, Az
The exterior orientation refers to a camera coordinate system
common to all CCD lines and is expressed for a given readout
cycle n as Xo, Yo, Zo, 9, €», x (Figure 1). The interior orientation
parameters xo, yo, c are defined in the image coordinate system,
three separate values exist for each line. The transformation
between the image coordinate system and the camera
coordinate system is given by Ax, Ay, Az, Aq, Ao, Ak, which
have been determined in the geometric calibration for each line
separately. M as well as D are rotation matrices, À is a scale
factor. The image coordinates are given by x and y, which are
derived automatically in this case via DIM.
MT(Ag, Aw, AK)
image coordinate system
MT(g, w, K, Aq, Au, AK)
camera
coordinate
system
object coordinate system
Figure 1: Coordinate systems (according to Kornus, 1999).
D'(g. t, K)
For the transformation from object space to image space as a
function of the image line (readout cycle) n the additional
condition (3) has to be applied.
x(n)ex (n, X,(), Y, (n), Z, (1), 0), eX), k(n))=0 GB)
847
This problem can be solved using the well known Newton-
method for the above zero-crossing detection where the
derivative x'(n;) is replaced by the pixelsize of the image.
n, — initial value for the image line :
oii (4)
n,, n, — x(n, )/ pixelsize i z 0,l,...
The principle of transforming a feature from one image to the
next is shown in Figure 2. The point P (extracted feature P") has
an estimated elevation Zp taken from the MOLA DTM, where
Az denotes the uncertainty of this value. This defines a range U,
L that is projected to the right image where it defines the search
space s.
C' C
MOLA
Figure 2: Principle of estimating matching location and search
space (according to Schenk, 1999).
After matching all overlapping images pairwise in all
combinations an undirected graph is generated. The nodes of
the graph are the point features, the edges are the matches
between them. This graph is divided into connected
components. The next step is the generation of point tuples,
whereas one point tuple is characterised by the property that not
more than one feature per image is admissible. The complexity
of this problem can grow exponentially. Instead of using tree
search or binary programming techniques a RANSAC (Random
Sample Consensus) procedure (Fischler, Bolles, 1981) is
applied. The method relies on the fact that the likelihood of
hitting a good configuration (correct tuple) by randomly
choosing a set of observations (features of the subgraph) is
large after a certain number of trials. The advantage of this
method is the high probability of getting a good point.
Including a geometric consistency check, the method also
eliminates blunders (Brand, Heipke, 1998).
From the start pyramid level (lowest resolution) to the so-called
intermediate level (medium resolution) feature based matching
(FBM) is carried out using the whole images. For the
processing of the HRSC imagery level 3 has been chosen as
starting level. Going down the image pyramid the image size
increases, as well as the number of extracted features. Besides
the heavily increasing computational time, the matching of the
complete images would result in too many tie points for the
camera orientation. Therefore the matching procedure is carried
out only for selected "image chips", starting at pyramid level 2.