s and 
s one 
jector 
| very 
rtions 
o not 
on of 
o the 
orrect 
nages 
| fre- 
ation 
onto 
ately 
cam- 
each 
by a 
rface 
f the 
areas 
bject 
:ulat- 
s one 
bject 
g the 
lange 
In is 
inical 
xd in 
on a 
more 
of all 
991). 
1 the 
lines 
able. 
f the 
this 
xture 
lated 
hing 
  
If the epipolar constraints of more than two images is used to obtain 
corresponding image details a fine matching procedure must follow. It is 
sufficient to use only two of the images for matching as all additional 
images are used to eliminate ambiguities. Again the least squares match- 
ing algorithm can be used or in case of a pattern of bright dots, the calcu- 
lation of the centre of gravity of the dot is also possible. 
6. PROJECTION OF A RANDOM PATTERN 
Least squares matching only gives good results if there is a texture of 
reasonable contrast on the object surface. If this is not the case an 
artificial pattern must be projected onto the object. This pattern should 
have a random structure to avoid ambiguities. High frequency random 
elements (such as single bright dots) are useful for the elimination of 
ambiguities and for good accuracy of least square matching. One the 
other hand a low frequency texture increases the radius of convergence of 
the matching procedure. Therefore, our artificial pattern consists of 
randomly distributed bright dots laid over a low frequency greyvalue 
pattern where the dots are located in the middle of the darkest areas of 
the grey pattern. 
7. DIGITAL SURFACE MODEL AND KNOWLEDGE BASE 
The digital surface model used during the derivation is a 2 1/2 dimension- 
al geometry where the direction of the elevation is roughly the same as 
the direction of the cameras. A transformation to a really 3 dimensional 
model must be done in a separate step if required. The main problems for 
automatic matching are caused by occlusions, geometric discontinuities of 
the surface, border lines of the area of interest and radiometric discontinu- 
ities such as shadows or dark areas of the object and areas without 
texture. Some of these problems can be eliminated during the segmenta- 
tion step (shadows, too dark areas). As mentioned earlier bright areas 
without texture can be avoided by projecting an artificial pattern on the 
object. 
A knowledge base can help to solve the ambiguity problems by checking 
the reliability of computed results or by avoiding areas where matching is 
not required or impossible. If a basic knowledge base is not available or 
of poor quality there must be a possibility for an operator interaction. But 
even the best image interpretation software and the most sophisticated 
expert systems are not able to calculate error free results. Although 
reliability checks are essential and a good knowledge base can help to 
detect matching errors, operator supervision will still be necessary. 
The quality of all checks depends mainly on the setting of appropriate 
thresholds. These various thresholds cannot be constant for the whole 
image area. An adaptive setting is the only way to obtain results as 
accurate as possible covering as much as possible of the area of interest 
with surface values. This means that the quality of the surface model 
varies from point to point depending on the local conditions of the images 
or the object. Therefore, the quality of the whole model cannot be descr- 
ibed by one single parameter or value. Only a digital accuracy model can 
describe the reliability and accuracy of a surface model in a sufficient 
way. Eventually it is very easy to eliminate or include points below or 
above a certain level of accuracy depending on the requirements of the 
current application. 
8. MEDICAL APPLICATIONS OF THE SURFACE MODEL 
The photogrammetric part of the program only provides the basic 
information for further computations. More or less complicated software 
must follow to process the requirements of the doctors. A module for a 3- 
dimensional graphic display of the surface is very important and it might 
be included into the photogrammetric part. Firstly, a picture of the surface 
is still one of the best checks for gross errors. Secondly, many of the 
medical applications are monitoring tasks of the shapes of human bodies 
or parts of them. A surface display may be a vector based wire frame 
model or preferably a pixel based greytone model. Other modules are 
editing programs where the current surface can be changed according to 
planned corrective measures. The calculations of differences and their 
display is necessary for monitoring corrective measures Or healing 
progress in time series. 
9. FINAL REMARKS 
Digital photogrammetry with images taken from CCD cameras can yield 
accurate results. If the process need not be in real time there are various 
possibility to improve the geometric accuracy by applying more sophisti- 
cated algorithms. Although conventional photogrammetric methods using 
photographic films are still more accurate there are many applications 
where classical photogrammetry is too slow if the result must be available 
minutes after data acquisition. Very often speed (even though not real 
time) is more important than the highest accuracy. In such cases CCD 
cameras are appropriate photogrammetric tools. With a knowledge base in 
the background photogrammetric systems become an important tool even 
for non-photogrammetrists like doctors. 
REFERENCES 
Beyer, H., 1987. Some Aspects on the Geometric Calibration of CCD- 
Cameras. In: Proceedings Intercommission Conference in Fast Processing 
of Photogrammetric Data, Interlaken, pp.68-81. 
Beyer, H., 1991. Photogrammetric On-Line Inspection for Car Crash 
Analysis - Results of a Pilot Project. In: Proceedings of the First 
Australian Photogrammetric Conference, Sydney, Vol.2,Nr.34. 
Gruen A., Baltsavias E., 1988. Geometrically Constraint Multiphoto 
Matching. Photogrammetric Engineering and Remote Sensing, Vol.54, 
No.5, pp.633-641. 
Kim C.NG. Alexander B.F., 1991. 3D Shape Measurement by Active 
Triangulation and Structured Lighting. In: Proceedings of the First 
Australian Photogrammetric Conference, Sydney, Vol.2.,Nr.33. 
Maas H-G., 1991. Digital Photogrammetry for Determination of Tracer 
Particle Coordinates in Turbulent Flow Research. Photogrammetric 
Engineering and Remote Sensing, Vol.57, No.12, pp.1593-1597. 
Stahs T.G, Wahl L.M.,1990: Fast and Robust Range Data Acquisition in a 
Low-Cost Environment. In: International Archives of Photogrammetry 
and Remote Sensing, Zurich, Vol.28, Part 5/1. 
Trinder, J.C., 1989. Precision of Digital Target Location. Photogrammet- 
ric Engineering and Remote Sensing, Vol.55, pp.883-886. 
Trinder J.C., Becek K., Donnelly B.E., 1991. Precision of Image Mat- 
ching. In: Proceedings of the First Australian Photogrammetric 
Conference, Sydney, Vol.2,Nr.43.