into account the present (e. g. computer) facilities and 
technical boundary conditions, such as e. E. the capabili- 
ties of the matching algorithms. 
The system for which the concept is developped consists 
of a Zeiss Planicomp C130 with an A900 of Hewlett Packard as 
host computer. The analytical plotter is extended by two 
CCD-cameras C1600 by Hamamatsu, which scan a small appr. 
5 x 5 mm2 patch in each image with a linear pixel size of 
appr. 20 um. The A900 is provided with this image informa- 
tion, the scanning of the two patches of 256 x 240 pixels 
taking about 1 sec.. 
1.1 Hierarchical Measurement Procedure 
A procedure which is able to measure a surface covered by 
a complete photogrammetric model in one step seems to be not 
available at present. As already pointed out in the intro- 
duction at least two levels of a hierarchy have to be rea- 
lized in a semiautomatic system based on theodolites or ana- 
lytical plotters with integrated digital cameras. 
But also the lower level, namely the automatic matching 
procedure working on the digitized images has to be struc- 
tured in a hierarchy. This is due to the limited pull-in 
range of practically all matching algorithms, which are able 
to handle large relative distortions between the image 
patches. The pull-in range, i. ^e. the allowable errors of 
the (approximate) starting values generally is in the order 
of 1/3 of the measured unit. For matching algorithms this 
means maximum shifts of 1/3 of the window size, maximum ro- 
tations or shears of 1/3 = 20 ? or scale factors between 0.7 
and 1.3. Reason for these limitations is the nonlinearity of 
the problem, if it is formulated in a parametric form, or 
the lack of invariance of most simlarity measures, e. E. the 
correlation coefficient, for evaluating the quality of the 
match. For wide or even superwide angle photographs much 
larger relief distortions may occur at the border of the 
image if the terrain is hilly leading to failures of a 
straight forward implementation of a matching algorithm. 
There are two ways out of this situation: 
- sequential procedures, where the already measured 
points form a basis for predicting adequate approximate 
values. For high performance a good mathematical model 
for the surface has to be available. 
- hierarchical procedures, where the matching is per- 
formed in a course-to-fine manner, the higher levels 
providing the approximate values for the lower levels. 
The matching algorithms in the different levels do not 
have to coincide, on the contrary they should be adap- 
- 618 —