When looking at the systems these comments seem 
to fit quite well. The systems developed by 
photogrametrists are more general in dealing with 
the redundant information from over-determined 
systems, even though the other systems may use the 
redundant information in some steps. The parts 
which are of special interest for redundant 
information is the system calibration and the point 
determination (data analysis). 
3.2 Self-Diagnosis and Quality Report 
A system operating over a time period must be able 
to control the quality of the output and to do the 
proper corrections during operation if necessary. A 
simple way of detecting errors in the output is e.g. to 
measure control points which are compared with 
their nominal values. If the detected error(s) is to be 
corrected, enough information must be provided by 
the system to locate, eliminate and update the error 
source. 
To be able to do a statistical error propagation 
through the whole process, from image acquisition to 
data analysis, the different parts must be 
encompassed in a statistical framework, where results 
from one level can be be used in the next. The error 
theory developed for photogrammetry is well suited 
for this task since it already covers the image 
acquisition part and the adjustment part of the data 
analysis. Two parts in the process are however less 
investigated: 
- data extraction 
- robust adjustment methods 
The data extraction methods used in image 
processing, e.g. edge and point detectors, are very 
seldomly producing statistical values of their 
performance. Methods used in a photogrammetric 
system should be able to produce this type of values 
to enable a correct statistical treatment of data, e.g. the 
Fórstner interest operator (Fórstner, 1987). A 
statistical propagation is also needed if a theoretical 
prediction of the results are to be done before 
implementation and installation. 
To make the process less sensitive to gross errors, the 
adjustment of redundant data may be treated with 
more robust methods than the normal equally 
weighted LS. The statistical properties of such 
methods are not always known or possible to directly 
put in to the normal statistical procedures. 
Comments The self-diagnostic capabilities of the 
systems are of very different nature depending of the 
degree of automization and application. Those of the 
systems which are manually supported rely mainly 
on the operator to detect errors. The more automated 
systems have the ability of detecting errors and in 
some cases to correct them. 
Two of systems developed by photogrammetrists uses 
control points to detect any changes in the system 
orientation and will automatically update the 
   
orientation parameters if needed. They also use 
several cameras to get an internal control of the point 
determination. 
The two systems not developed by photo- 
grammetrists have other ways of detecting errors in 
the system calibration, e.g. known distances, but are 
not able to correct it without operator assistance. 
As mentioned in 3.1 the different view on how to 
describe the precision for the systems is valid also for 
the self-diagnostics and quality reports. The error 
theory which is used in the traditional photo- 
grammetric systems require a statistical model for all 
the different steps to enable an error propagation. The 
other approach is to empirically estimate the accuracy 
of the system and use these values without the 
statistical background. This is a fast and compu- 
tationally easy method and also easy to understand 
for the non-specialists who are to use the systems. 
3.3 Task Flexibility 
The third criterion implies that photogrammetric 
systems are fairly general systems with respect to 3D 
object reconstruction. It may be argued that dedicated 
systems and even single camera systems with 2D 
capabilities might in some cases should be regarded as 
photogrammetric as well, as long as the extracted 
information from the images is metric. 
Comments The generality of the 3D calcualations are 
partly depending on the type of measurments the 
system is able to do. Grey-level based data extraction, 
used by the two systems developed by photo- 
grammetrists, is in principle more general than target 
measurements, but many other factors, like sampling 
speed and data analysis, should also be considered. 
4. REAL-TIME SYSTEMS - THE SYSTEM PARTS 
Even though photogrammetric systems may differ 
between each other in many respects, they are usually 
having the same basic components (fig 1). 
Different tasks will certainly put different restrictions 
on the time constraints for the systems. Some 
applications have the hardest constraints on the 
image acquisition part, e.g. high speed motion 
analysis systems, where the extractions and analysis 
of data may not be completed or even started between 
the acquisition of two image frames. Other, more 
quality control or robot oriented tasks, may need to 
perform all steps in sequence in order to be able to 
make a decision in the time constrained cycle. 
In the following section the different parts of the 
photogrammetric system are discussed and the four 
systems are briefly described in this context.