am, 
nic 
m- 
us 
ral 
the 
Iti- 
ike 
er 
ne 
its 
9) 
ite 
ed 
ic 
hg 
to 
Je 
je 
These steps are performed from level to level of an 
image pyramid in order to minimize manual interaction 
and to make the process more robust. 
Iterative least squares techniques are used in photo- 
grammetric adjustment based mainly on collinearity 
equations for conjugate and ground control points. 
Some other equations are used for the introduction of 
known platform flight path and attitude information and 
for the estimation of their constant and higher order 
biases. Interior orientation parameters can be esti- 
mated, too. 
2. Image matching 
2.1 Interest operator 
In high resolution imagery like that from the airborne 
model of the MEOSS camera (about 2 . 2m? pixel size) 
large homogeneous areas like fields, meadows and 
water bodies appear where good patterns for image 
correlation cannot be extracted. Therefore, for auto- 
matic image matching, an interest operator has to 
automatically select patterns which are well suited for 
digital image correlation. Our interest operator is 
designed along the lines of the so-called "Fórstner 
operator” developed at Stuttgart university < Ref. 8>. 
The weighted centre of gravity is calculated for each 
window of a given size by least squares technique with 
appropriate weighting by first order derivatives in line 
and column directions. From the error ellipse of this 
least squares adjustment two parameters, roundness 
and size, are extracted. Windows and their resulting 
points are said to be promising for image correlation if 
the size is small and the roundness is beyond a given 
threshold so that the points are well defined in terms 
of multidirectional edge information contents of the 
surrounding window. 
We found that the sizes of the error ellipses are corre- 
lated with the variances of the grey values of the win- 
dows (about -0.5 normalized correlation coefficient). 
Furthermore, it was found through many tests that win- 
dows - even if they meet the requirements on round- 
ness and size described in «Ref. 8^ - should have a 
variance beyond some threshold. Thus, for the user 
interface the threshold for size is replaced by a thresh- 
old for variance. This variance thresholds can be esti- 
mated from the local image variances much better than 
the median value of the sizes of the error ellipses 
mentioned in «Ref. 8». For our airborne imagery 
thresholds between 25.0 and 64.0 for the variance gave 
reasonable distributions of the points found by the 
interest operator. 
Normally, a promising point is located more than once 
by the operator (with respect to different windows). This 
multiplicity is reflected in our procedure via a count 
only. Huge numbers of points are registered often by 
this procedure. Thus, it is found a good strategy to 
replace a set of points which are lying near to each 
other by one prominent point. The selection of the 
prominent point is based on the following parameters, 
given in descending order of priority: 
69 
multiplicity of point 
variance of window 
roundness of error ellipse 
size of error ellipse 
2.2 Search window selection 
The interest operator is used to locate good patterns in 
the nadir looking sensor's scene. For area based 
matching corresponding search areas have to be 
extracted from the scenes of the other looking 
directions. This is achieved by computing a local affine 
transformation between the stereo partners. The six 
parameters of the affine transformation are computed 
by least squares adjustment using already known con- 
jugate points. Currently, at the lowest level of resolution 
of the image pyramid manually found conjugate points 
are used as input to this process. This could be further 
automated by using some a priori knowledge about the 
sensor geometry (e.g. stereo baselength in number of 
pixels) and interest operator results for all stereo part- 
ners. 
The results of the local affine transformation are 
accepted only if the rms-error for the input conjugate 
points is less than half the possible shift of the pattern 
area within the search area. Thus, coarse errors in the 
position of the search windows are avoided. More trials 
are made by changing the number of input points 
and/or the the weighting scheme. If none of the adjust- 
ments is accepted no matching process for this point 
and stereo pair is initiated. The number and distribution 
of these affine transformation failures indicate to the 
user in what regions of the imagery the point densities 
for computing search area positions are not sufficient. 
2.3 Image correlation with pixel accuracy 
A matrix of normalized correlation coefficients is com- 
puted for given pattern and search areas by shifting the 
pattern pixel by pixel (in column and line directions) 
over the search area. The maximum of the correlation 
coefficients defines the location of that pixel in the 
search area which corresponds to the centre pixel of 
the pattern area. A quality figure is defined which mea- 
sures the uniqueness and the relative steepness of the 
peak in the matrix of correlation coefficients (for defi- 
nition see «Ref. 47» ). 
Strict acceptance rules are applied to the results 
. of each individual correlation process between a 
stereo pair and 
e of 3 combined correlation processes if 3 stereo 
pairs are available (in the zone of threefold stere- 
oscopic coverage). 
At the level of a single correlation the following condi- 
tions have to be met before acceptance of the results: 
* maximum of correlation coefficients and quality 
figure have to be beyond given thresholds (in case 
of our data material 0.5 and 0.2 were taken for 
correlation coefficient and quality figure, respec- 
tively) 
® the maximum of the correlation coefficient must not 
lie on the border of the matrix of correlation coef- 
ficients (the width of the border is normally set to 
1 column and 1 line)