o different im- 
RE or PARAL- 
| same 3D sur- 
examining rays 
n and by pass- 
cts. The rays 
ction. On ideal 
bonding object 
the 3D-space. 
and inaccurate 
> not intersect. 
e rays is calcu- 
ot correspond- 
f pairs of ver- 
Additionally, a 
] out (e.g. re- 
t not be under 
spond, the ob- 
If objects 
at the surface 
ithin a certain 
a way that the 
object ROOF is 
assumed to be 
sis is complete 
scription of the 
reprocessing is 
) is transferred 
into a sequence of n, binary images (Fig. 4b) by n, 
thresholds. Possessing previous knowledge of the inten- 
sities of objects in greylevel images, a certain distribu- 
tion of thresholds can be set up. Without this knowl- 
edge thresholds are distributed equidistantly between 
the minimal and maximal greylevel in the image. The 
number of thresholds n; is a process parameter (which 
is usually chosen n, > 8). 
(2) Then the contour lines of the segments in each bi- 
nary image are detected (Fig. 4c). Considering the im- 
age greylevels to be altitudes of a mountain, we obtain 
the lines of equal altitudes of this mountain by segmen- 
tation and contour tracking. In areas of high gradients 
we will receive many altitude lines, in areas of low gra- 
dients few altitude lines. 
(3) In the next step of preprocessing the contour lines 
are approximated by short straight lines (Fig. 4d). This 
process of approximation is done by applying a dynamic 
split algorithm, which works similar to the algorithm of 
Ramer [1972]. In order to carry out such an approxi- 
mation the contour line must have a minimum length. 
A termination criteria is given by a minimum quality of 
approximation and a minimal length for line segments. 
(4) In the last step the short line segments of the 
contour sequence are grouped together to long lines. 
The long lines are stored as primitive objects LINE 
(Fig. 5, L, left) in the blackboard. They are at- 
tributed with length, orientation, coordinates, original 
image (left/right) and an assessment. The assessment 
is deduced from the quality of approximation. 
  
  
  
  
  
  
  
  
Fig. 4: Preprocessing (section of left image). 
a) greylevel image, b) binary image sequence, 
C) contour image sequence, d) short lines 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
left right 
Fig. 5: 3D-Analysis. Interim results displayed by 
different sets of objects. 
835