106 
  
1 
I 2 7 
T ans 
STR 5 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Figure 2: Partition of an image (left) and states of the cognitive model (right) 
4. DETECTION OF ATTENTION FIELDS 
The attention control in MOSAIK (Sobottka, 1994) aims the early detection of new objects in the scene. It 
is based on the idea firstly to segment the image into regions R;, à — 1,...,n of uniform motion and then to 
decide for every region if it is relevant to the task or not. This decision is done based on perceptual grouping 
strategies. The determined regions of interest, so-called attention fields, are the result of the attention control 
and are denoted by A;, j = 1,...,m (see figure 3, left). This kind of problem reduction make it possible to 
process also complex scenes in a very short computing time. As input the attention control receives the results 
of edge detection in MOSAIK. 
image t t 
| d peter 8. 
senuane le t-3 J 
ms == A 
motion analysis | [eo | ; D 
= 
LT = 
YI 
  
  
  
Y 
y 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
R, R5 Ra Rn (a) max. displacement (b) temp. continuity 
t : t 
l = ----- = 
; i DÀ... IR— 91 pe 70707 3 
perceptual grouping Let = T ue = + 
ef TT | | =| = 
| — EL RR 
A A2 Anil TA po 
mi m (c) consistency (d) similarity 
  
  
  
  
  
  
  
  
  
  
Figure 3: Problem reduction in the approach for attention control (left) and constraints to the correspondence 
problem (right) 
We determine motion by using line correspondence. The method includes three steps: Line detection, line 
matching and segmentation of the displacement vector field. 
We choose line segments as interesting features because they are representative and stable (Xie et al., 1993). 
Another advantage is that in particular horizontal or vertical line segments characterize road vehicles in a 
significant manner. In practise, line segments are often not perfectly horizontal and vertical. Therefore, we 
consider a tolerance threshold of ten degrees. The detection of those near-horizontal and near-vertical line 
segments is done by heuristic search strategies. 
In the second step corresponding line segments have to be determined. To simplify the correspondence problem 
we make use of the maximal possible displacement between two consecutive frames, continuity of motion, 
consistency of line matching and similarity in line parameters (Fig. 3, right). 
The maximal displacement (a) can be used to constrain the search space of corresponding line segments in 
consecutive frames. It can be computed based on the maximal velocity of vehicles. The temporal continuity 
constraint (b) specifies, that the motion of objects doesn't change abruptly. So a prediction of the position of a 
line segment in the next frame is possible. The consistency of line matching (c) requires that one line segment 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences’, Zurich, March 22-24 1995