Fig. 2 Blue band of a scanned colour aerial photograph, pixelsize 1.60 m., 
representing the present situation. The considered road segment from the 
database is superimposed on the image in white. 
This introduces the difficult task of hypotheses 
hierarchy. That means: What should be the sequence 
of the hypotheses to be tested. 
To test the last hypotheses, we need the result of the 
first two. Real-world knowledge can be used to 
determine the hypotheses hierarchy. 
  
Using this knowledge the first hypothesis to be tested 
should be whether the road segment is still present. 
Then the problem is faced of choosing appropriate 
segmentation techniques to test the hypothesis. This 
choice should at least depend on the following criteria: 
- The segmentation techniques should produce 
relevant parameters, for example the width of the 
road, that can be employed during hypothesis in a 
subsequent stage of the road hypotheses hierarchy; 
- They should agree with characteristics of the data, 
for example the noise level and pixel size (satellite 
images require other techniques than large scale 
aerial images); 
- They should agree with the type of object, for 
example for testing the presence of an hypothesized 
fly-over, we could use a technique that detects sha- 
dows. 
To be able to make this choice, we need knowledge 
about operating characteristics of segmentation techni- 
ques, knowledge about the image data and real-world 
knowledge. 
  
Fig.3 Region of interest (ROI) defined by the road segment. The ROI was extracted from the image by resampling