Input Data
|
| Feature Extraction |
|
| Feature Based Matching|
Y
Least Squares Matching
with Region Growing
Y
Tracking
|
Conjugate Points
Figure 2: Work-flow of the matching concept
path, i.e. the relative accuracy will be very good. The ab-
solute accuracy, however, will be in the order of 1 km and
2' respectively. Consequently, it will be sufficient to use
the orbital data solely, if only images of one single strip
are matched. In order to connect images of different strips
(orbital arcs), tie points should be available.
2.2 Feature Extraction
Feature extraction is performed on the highest level of the
image pyramid, i.e. on the level with the lowest resolution.
It is carried out for each image independently. Currently
the Moravec (1977) interest operator is implemented. This
operator do not really locate point features, but a kind of
interest areas (Fôrstner and Gülch 1987). Within these ar-
eas significant differences of the gray values exist in each of
the four main directions. For this reason conjugate points
may be found using an intensity based procedure, e.g. gray
value correlation. The position of the areas is calculated
with an accuracy of one pixel only. Hence the Moravec
operator is not suitable for accurate matching but can
provide good initial values for a more accurate method
(e.g. least squares matching).
2.3 Feature Based Matching
For each interest point of the reference image, the corre-
sponding interest point of the second image is ascertained
by calculating the correlation between the image windows
in the surroundings of the two interest points. If the corre-
lation coefficient lies above a certain user-defined thresh-
old, the two points are considered as conjugate points. The
choice of this threshold influences number and reliability
of the conjugate points. In general, a high threshold value
(e.g. 0.95) will provide highly reliable points whereas a
lower value (e.g. 0.7) leads to a dense but less reliable
point distribution.
This matching procedure would be very time consuming
and susceptible to ambiguities if each interest point of the
second image would be considered as a candidate for a
942
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
conjugate point. For that reason, for each interest point
in the reference image some coarse information about its
estimated position in the other image is utilized to build
up search areas. This geometric information is taken ei-
ther from the control points or from the orbital data, from
which affine transformation parameters are calculated. As
the scale may vary within one image due to the elliptic
Mars'96 orbit, it is necessary to partition the image into
transformation cells, each having its own set of transfor-
mation parameters.
From feature based matching a set of conjugate points is
provided. If the acceptance threshold is chosen high, the
resulting pairs of conjugate points are reliable but not very
dense. The following method is introduced to receive a
dense distribution of conjugate points.
2.4 Least Squares Matching
This method consists of two major steps: First, the con-
jugate points are checked and their accuracy is improved
by means of least squares matching (Ackermann 1983).
Second, these re-measured points serve as seed points for
region growing (Otto and Chau 1989).
The number of resulting points mainly depends on the
raster spacing of the region growing and the pyramid level
on which it is performed. The raster spacing should not
be too large, especially in mountainous regions in order to
get good approximations for the next points. The pyramid
level should be chosen dependent on the complexity of the
image data, e.g. if there are only structures of only a few
pixels width which possibly disappears on higher pyramid
levels, a lower start level have to be chosen.
To speed up the matching procedure, least squares match-
ing with region growing normally is performed only once
on a higher pyramid level (e.g. level 5, where one pixel
consist of 32 - 32 pixel of the original image).
2.5 Tracking
If the parameters (step size, pyramid level) are chosen ap-
propriately, enough conjugate points are generated by least
squares matching. To improve the accuracy, the conjugate
points are projected on the next lower pyramid level and
remeasured using least squares matching. This step is re-
peated until the lowest level is reached. The number of
points decreases because of
e wrongly matched points which are eliminated on dif-
ferent resolution levels
e the non-linear change of texture within the image
pyramid
3 PRACTICAL TESTS
In this section we present results of the new matching
approach. To this end, three different image sets have
been selected. Two of them were taken by 3-line scanners
whereas the third was acquired by a frame camera. For all
image sets we used six manually measured tie points per
image as initial information.
Fi
In
Sc
hz
si:
17
fo
ar
le
in
of
de
In
by
tr