formation. These features are also good if we consider 
the definition of initial values for LSQ-estimation. 
Another curve like a b-spline is an ideal feature to 
depict many natural curves, but the problem of 
parameterisation and getting reasonable initial values 
restrict the use of it. 
The Hough transformation is a widely known method 
for finding points which belong to a curve with a 
predetermined feature type. The disadvantage of the 
method has been its great computer consumption in 
time and memory. Many variations have been de- 
signed for decreasing the computational complexity of 
the method as well as improving the accuracy of the 
algorithm. The algorithms have been divided into 
standard algorithms and stochastic Hough trans- 
formations. One approach is to use statistical analysis 
to determine correct parameters in transformation 
space. This approach has been successfully investi- 
gated by J. Kittler, J. Illingworth, J. Princen and 
H.K. Yuen’. The disadvantage of this kind of ap- 
proach is the complexity of the computation. The accu- 
racy of the method as well as the reliability are 
remarkable. Another fascinating algorithm, proposed 
by L. Xu and E. Oja’, is the Random Hough 
Transformation which fulfills both the expectation of 
accuracy and saving of storage space. 
3.1 RHT Randomized Hough transformation 
The Randomized Hough transformation RHT belongs 
to the category of probabilistic Hough transformations. 
The idea of the method comes from neural computing. 
By inspirations of Kohonen map, the algorithm uses 
converging mapping to determine the Hough par- 
ameters. The whole method relies on random sam- 
pling, converging mapping and stepwise implemen- 
tation of accumulation. Xu and Oja have introduced 
the algorithm for usage with different kinds of score 
storage structures, but the dynamic list structure 
appears to be the best at the point of view of storage 
space. 
The procedure goes by alternating the converging 
mapping and accumulation periods. The procedure 
requires predetermined values to terminate the 
process. The number of maximum trials for verifying a 
feature existence must be heuristically set. This of 
course depends on the case. With suitable values the 
algorithm finds quite accurately all points belonging to 
each feature and is much faster than other comparable 
methods. The processing scheme of RHT-algorithm is 
depicted below. 
       
222 
Hough Transformation 
  
  
  
  
  
Figure 1. Procedure scheme of RHT. 
The algorithm was applied for edges extracted with 
Canny operator, Figure 2. The edges falling into sub- 
pixel gap in Hough space are depicted in Figure 3. 
Figure 2. Results of the edge detection. Canny oper- 
ator. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996 
  
In this cha 
image spa 
belonging t 
recognition 
3D linear 
matching t 
dence probl 
When mea 
displaceme 
tive frames 
parameters 
feature, bo 
average str 
case of lin 
ending poir 
Matching o 
be a ambig 
all combin 
similarity 
between fe 
Often som 
correlation 
measures 
determined 
Finding th 
the first fi 
matching, : 
method wi 
probabilisti 
that the no 
node. Rela 
result can | 
nodes are v 
A problem 
a such case 
set fora s 
Also geom: 
can stabili 
camera mo