2.4 Data reduction 
Before moving to a higher level of processing, we try to get 
segments likely to be grouped and linked in significant 
structures. This means to connect strictly collinear very 
close segments in one new longer segment: in other words 
we simply try to make up for narrow gaps arisen along an 
edge. In all cases we processed, this amounted to a 15% 
reduction of the number of lines. After and only after this 
stage, we remove all short segments(less than 2 pixel long) 
since they won't play a role in the subsequent stages. Out of 
the original number of segments, 80% are discarded at the 
end of this stage (see Figure 4). 
3. PERCEPTUAL ORGANIZATION 
On the way to image interpretation, we need to move from a 
description based on gray values to a more abstract level, to 
identify structures. These may be termed as collections of 
elements (lines or regions) which the visual human system 
perceives as connected or interrelated, even without any a 
priori knowledge of their contents: this process is called 
perceptual grouping. We look for relations which should be 
least sensitive to changes to the standpoint and with small 
probability to arise in the image by chance (ie. a 
radiometric edge will truly represent a physical edge), for 
instance collinearity, proximity, closure, parallelism... 
(Lowe, 1985; Sarkar & Bayer, 1993). 
We used proximity to reduce the search space, since 
features which are far apart in the image are not likely to 
share significant connections. A search window (see Figure 
5) is build around the gravity center of each segment using 
the value Dmax of the maximum width of the road classes 
for that image scale. 
Dmax 
  
€ 
e 
i, 
or o E eer 
Figure 5. The search window 
Within the search window, we look for collinearity and 
anticollinearity (that is for segments sharing the same 
direction, but with opposite gradient orientation), 
cocurvilinearity (and anti-cocurvilinearity), parallelism (and 
anti-parallelism), junctions (see Figure 6). To correctly 
label the relations, we must decide to what extent the actual 
—l— —— Collineaity —+— —— 
Amin Edge gradient direction T 
pov 577 EDU 
eee a 
Figure 6. The basic relationship between line segments 
204 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
features match an ideal model (e.g. two segments in reality 
will never be parallel in a strict mathematical sense). This is 
simply achieved by setting up some threshold values for 
distances Ar and Av between the end points and differences 
in direction At for each pair of segments (see Figure 7). 
Parallel to this process, we classify the attributes of features 
and relations. Figure 8 shows the attributes recorded for 
each relation. 
We have now completed a relational description of our 
primitives, that is, of the line segments. 
  
Figure 7. The elements used to define the relation 
between a pair of segments 
  
  
Sab-tSba.5. nel 
S= 
2 
_Dab + Dba 
2 
d 
  
  
  
Parallelism 
Figure 8. Attributes of relations 
    
  
  
  
  
  
   
   
  
   
   
  
  
  
   
   
   
  
   
   
  
   
   
  
   
      
   
  
    
    
   
  
  
  
   
  
  
   
   
   
   
   
   
   
   
    
Figure 
Our reco: 
assumptiot 
e roads] 
Babu, 198 
e roads: 
e roads: 
In our ex 
homogene 
often. 
We proce 
antiparalli 
recursivel 
that is, a 
primitives 
higher le 
continuity 
curvilinea 
At th 
antiparall 
more tha 
the grou; 
network 
between 
Ideally, i 
this proc: 
networks 
algorithn