Operations can be automated entirely or in 
- part. An example of partial automation, or 
computer assisted operations is when the 
operator sets an approximate position and the 
final position is fixed automatically. This 
allows the computer to perform the difficult 
operation and usually to achieve maximum 
precision. An example is the line extraction 
routine in the ISTAR VUE3D system in which 
the operator roughly indicates the position of a 
line and the computer finds the exact line of 
the feature using a dynamic programming 
technique developed by Mâitre and Wu 
(1989). The stereo matching of two images to 
ensure correct height when the operator 
follows a feature is another example which 
was introduced in the Kern DSP1 and is now 
incorporated into a number of systems. 
In the production systems automation does not 
tend to be introduced until a robust algorithm 
has been developed and proven. Thus most 
systems still rely on the operator to carry out 
the inner, relative and absolute orientation with 
support from image processing routines such 
as zooming. 
Software for the production of DEMs is now is 
use for production and DEMs can be produced 
automatically but still need to be edited for 
blunders and missing areas. 
3.3 Research directions 
With the exception of the DEM extraction, no 
commercial systems has a proven automated 
component. It is evident from the literature 
that there are a number of areas where 
automation is seen as having a potential, 
either in the near or distant future. These areas 
are: 
* identification of ground control points; 
* Speeding up the orientation process 
necessary to determine the exterior 
orientation of the images; 
* feature extraction; 
* change detection. 
3.3.1 Automatic GCP extraction 
There is a significant amount of work going on 
to reduce the dependence on ground control 
points (GCPs) for absolute orientation and 
georeferencing procedures. 
The identification and selection of GCPs for 
geometric processing of digital images is 
usually a time-consuming and expensive 
process although discussion with producers 
indicates that this is a small proportion of the 
total cost of producing image maps. Automatic 
340 
matching with earlier processed images is a 
significant help in removing this bottle-neck. 
This has been discussed and implemented for a 
number of years, for example Benny (1981). 
A major problem to be addressed is the 
accuracy and distribution of extracted points. 
The problem of map-image matching is much 
more difficult as work by UCL has shown 
(Stevens et al 1988). 
Building on earlier work (Schenk et. al. 1991) 
on automatic tie-point determination for the 
orientation of digital photographs, Toth and 
Schenk (1992) have described a method of 
automatically registering images by matching 
extracted edges and determining identical 
points. 
3.32 Automation within the orientation 
process 
In order to determine the orientation elements 
it is necessary to establish the calibration 
parameters of the sensor and to fix relative 
orientation and absolute orientation using 
ground reference points. The points required 
for calibration and relative orientation 
(conjugate points) can already be determined 
automatically. The fiducial points and the 
conjugate image coordinates can be derived 
from stereo matching. 
Stokes (1988) developed a fully automated 
procedure to identify and measure fiducial 
marks. Fiducial marks are generally different 
in different cameras. However, they usually 
have a well defined appearance and occupy an 
extended area devoid of other information. 
Their degree of symmetry is high and their 
approximate location in the image known in 
advance. They can therefore easily be 
identified using template matching and then 
localised with centre of gravity methods. 
Haala et al (1993) have described work 
leading to full automation of the conjugate 
point problem starting with an interest operator 
and using an image pyramid to refine the 
match. 
A number of organisations are working with 
image registration systems which consider 
whole images (Lee et al, 1993) or layers in 
images such as roads (ENST in Paris using 
techniques described by Maitre and Wu, 
1989). 
Schickler (1992) developed a system for 
automating the exterior orientation of a single 
image. It is based on control points which 
consist of a list of straight 3D-line segments, 
whose 3D-coordinates are known in a object 
centred coordinate system, and which mostly 
  
repre 
valuc 
to in 
the ir 
segir 
imag 
avail 
resec 
segim 
preci 
This 
of di: 
phot 
more 
point 
space 
plan: 
their 
for i 
techi 
essel 
than 
orde 
extrz 
featu 
must 
WOrk 
geon 
perf 
phot 
Haal 
relat 
desc 
insta 
back 
struc 
obtai 
colo 
linea 
the | 
map: 
been 
Syst 
3.3.3 
Feat 
inter 
defir 
robu 
McK 
majc 
varia 
othe: 
Syste 
comi 
rese: 
the 
relat 
with 
mult