The latter condition is very helpful in regions with 
repetitive patterns. 
For a full conjugate point where three stereo pairs are 
available the conditions are as follows: 
* for all three correlations the above mentioned cri- 
teria for single correlations have to be fulfilled 
* for the checking correlation between the back- 
ward/forward stereo pair the absolute value of the 
displacement vector resulting from the correlation 
must not exceed a given limit (normally set to 1 
column and 1 line). 
Pattern sizes have been 7 . 7 pixels for the 3 lower lev- 
els of the image pyramid and 9.9 for the two higher 
levels; accordingly, search area sizes of 15.15 and 
21-21 were taken to cope with the increase in paral- 
laxes at the higher levels. This is depending also on the 
density of the conjugate points found for the previous 
level of the image pyramid. Of course, high densities 
will help to reduce the search areas. 
2.4 Subpixel accuracy 
Local least squares matching techniques (LSM) 
described in <Ref. 9> are used to refine the results 
of the previously described operator to subpixel level. 
The radius of convergence of LSM is a few pixels only. 
Thus, it has to be preceded by a matching operator at 
pixel level. 
The parameters of two local transformations between 
the stereo partners - two parameters of a radiometric 
transformation (brightness and contrast) and six 
parameters of an affine transformation for obtaining the 
subpixel positional information - are estimated by iter- 
ative least squares adjustment. The observations are 
the differences of the grey values of the original scene 
of the nadir looking sensor and the transformed sub- 
scene of the stereo partner. It was found that the con- 
vergence is bad for windows less than 11*11 pixels for 
the given data material (in <Ref. 9> 16*16 is recom- 
mended as smallest window size - but for frame camera 
imagery). The smoothing of the grey values before LSM 
mentioned in <Ref. 9> is not realized in our software 
because very good initial values are provided by cor- 
relation at pixel level. 
For most windows the convergence is within 2 to 5 
iterations. In case of non-convergence the location of 
the maximum of a parabolic fit to the neighbourhood of 
the maximum of the correlation’ coefficients is taken as 
a substitute. This also defines the initial values taken 
for the LSM. 
Experience showed that the parabolic fit maxima will 
often differ much from the LSM results. Convergence 
of LSM was achieved in about 90% of the cases. 
2.5 Image pyramid 
Already our short strips of stereo imagery of MEOSS 
type are a massive amount of bytes on disk or subareas 
fitting to normal display screens. Additionally, dis- 
tortions introduced by the terrain and the attitude vari- 
ations of the aircraft are often very large even locally 
(even more than 100 pixels). Thus, manual starting of 
locating conjugate points would be a very tedious task. 
These problems are much reduced by introducing a 
resolution pyramid. If we use factor 16 for the coarsest 
resolution (in both line and column direction) this 
results in small scenes fitting to modern display 
screens. Furthermore, as parallaxes are reduced by the 
same factor, the human operator will be able to quickly 
measure the small set of conjugate points required for 
starting the search area selection. All further steps of 
interest operator and image correlation work automat- 
ically through the image pyramid up to the finest level 
of resolution. 
The quality of the final results is profiting much from the 
fact that from one level of resolution to the next the 
increase in distortions is relatively small. Of course, 
one has to pay for this by an increase in computer time 
and disk storage (though a full pyramid of five levels 
results in a storage increase by a factor 1.33 only). 
3. Photogrammetric combined point determination 
3.1 Basic equations 
The iterative least squares adjustment for computing 
the ground coordinates of the conjugate points and 
improved values of the exterior orientation parameters 
is based on the following types of equations: 
© collinearity equations connecting image coordi- 
nates and ground coordinates of the conjugate 
points with the exterior orientation of the camera 
at certain orientation images 
* observation equations for ground control points 
* observation equations for the parameters of interi- 
or orientation of the camera (position of principal 
point and focal length) 
* observation equations for exterior orientation 
parameters including constant and higher order 
biases 
* a second order Gauss-Markov process for the 
parameters of exterior orientation (this is meant 
primarily for bridging large gaps in the distribution 
of conjugate points caused for example by large 
homogeneous areas, water bodies and clouds). 
The exterior orientation is calculated for a set of orien- 
tation images. These may be selected with regular or 
irregular spacing in time along the orbit. Currently, lin- 
ear interpolation is used to obtain the exterior orien- 
tation for each line of the scanner imagery < Ref. 
2,10>. Thus 12 parameters of exterior orientation 
enter into the two collinearity equations derived for one 
imaging ray. 
3.2 Input data 
The input to the photogrammetric adjustment consists 
of: 
* the conjugate points found by semi-automatic 
image matching (these are the most precise meas- 
urements) 
* ground control points: these are conjugate points 
which are also identified on maps; topographic 
maps are used to extract GauB-Kriiger coordinates 
and heights (this is a very tedious and time con- 
suming manual work) 
e. initial values for the exterior orientation at the ori- 
entation images 
geometric calibration data for the camera 
® weights for all the error equations. 
41 
The 
sel 
act 
be 
ler 
pla 
in 
The 
tor 
giv 
pyt 
Fo 
33 
Wa 
the