PP A 
Real-time, time constrained and video-rate are terms 
which are related and partly overlapping each other. 
By video-rate is here meant a standard video imaging 
system, generating images at 25/50 Hz. Many systems 
uses its own image acquisition speed, which here are 
related as being either higher or lower than the 
standard video-rate. Real-time in machine and robot 
vision are often implicitly meant to be the standard 
video-rate. A more general definition of real-time is 
time-constrained, giving a limited time in which the 
task must be solved. In this paper real-time systems 
are equivalent to time constrained systems. 
By Calibration is, in the photogrammetric society, 
mostly meant the determination of the inner 
orientation of a camera. In machine vision the term 
often stands for the determination of the outer 
orientation parameters as well. In this paper System 
Calibration is referred to as the determination of the 
absolute orientation and the inner orientation if they 
is determined simultaneously. If the inner 
orientation is determined separately, this is referred 
to as Camera Calibration. The term system calibration 
is chosen in favour of outer orientation or absolute 
orientation since it is more relevant when talking 
about an industrial installation. 
3. A PHOTOGRAMMETRIC SYSTEM 
As the title of this article indicates, the primal interest 
is in the developing process of the close-range 
systems, not so much the actual performance of the 
systems themselves. To be regarded as photo- 
grammetric, a close-range system must however 
meet certain criteria. One suggestion for these criteria 
are given by (Grün, 1991): 
- Potential for high precision and reliability 
(redundant sensor data) 
- Capability of self-diagnosis (quality report) 
- Task flexibility with respect to 3-D object 
reconstruction functions 
This 'definition' of a photogrammetric system may 
be valid within our own society, while in computer 
vision the term 'photogrammetry' usually stands for 
the various orientation procedures of stereo images 
which here is related only to the third criterion. The 
definition implies of course that a system developed 
by a photogrammetrist may be said to be non- 
photogrammetric, while a system developed by a 
machine vision engineer may be seen as 
photogrammetric in our eyes. 
When designing a real-time measuring system, all 
three criteria will by their nature be in conflict with 
the time constraints, since the time complexity of the 
computations are high for each of them. 
3.1 High Precision and Reliability 
High precision is possible to achieve with the 
methods available in data extraction and data analysis 
   
   
  
  
  
  
  
  
  
  
  
  
  
  
  
   
   
  
  
  
    
  
   
  
  
  
  
  
   
  
  
  
   
  
  
  
   
  
  
   
  
  
  
  
   
  
  
  
  
  
   
  
   
  
  
  
  
   
  
  
    
   
  
    
(e.g. Haggren, 1990). Several systems aiming at high 
precision reach results which are as good, or better, 
than a human operator in cases where the targets are 
well defined. 
The reliability is a more delicate matter since it 
touches the part of a system which is harder to 
describe in statistical figures: its insensitivity, or 
robustness, against erroneous data or model outliers. 
In a manually operated system, gross errors are rare 
and fairly simple methods can be used to locate them. 
When a process is automated or semi-automated, as 
in the case of real-time measuring systems, the need 
for more robust methods becomes more obvious. The 
robustness should be incorporated in all the parts of 
the measuring process, to ensure that single or 
groups of erroneous data do not influence the final 
output. 
Two aspects of robustness are of major concern 
(Fórstner, 1987): 
- Robustness of design 
- Robustness of estimation 
The robustness of design is concerned with the ability 
to test the models with respect to model errors and 
with the sensitivity of the result to errors. The Least 
Squares, LS, techniques together with statistical 
analysis are the main tools. 
Robustness of estimation is concerned with 
optimization procedures which eliminates or reduces 
the effect of model errors. Other types of estimators 
which are more robust against model errors than the 
LS have been developed, e.g. Least Median Squares 
(Rousseeuw, 1987) and Minimum Description Length 
(Axelsson, 1992). These estimators, which can handle 
up to 50% of erroneous data, all lack an analytical 
solution. Instead, a systematic or random search must 
be used for finding the solution. This makes the 
methods computationally very complex. If the 
number of parameters are very high, as e.g. in a 
bundle adjustment these methods are not suitable. 
For other applications, like e.g. relative orientation or 
orientation of a single camera, the methods should be 
considered. 
Comments None of the illustrating systems uses the 
second type, robustness of estimation. These methods 
are fairly new and the knowledge of them limited 
outside the statistical research environments. We 
believe that these method will play an important role 
in future systems, both in the extraction of image 
features and in orientation procedures. 
The general view on the precision concept and if 
there are differences depending on the background 
was formulated by an electrical engineer as 
"..photogrammetrists think of precision in the 
cameras, images and all the different steps. We only 
relate to the deviations from a known reference 
object..". From the 'photogrammetric' side one 
person said that "..redundant observations are not 
fully utilized in non-photogrammetrical systems...".