The most important result of that work can be summarized 
as follows. The risk to fail with an area based matching 
technique increases in structures with repetition, because 
the number of correlation peaks is higher than one. For pre- 
cise localization of a structure it is advantageous to have a 
steeply descending autocorrelation function. The gradients 
listed in figure 4 indicate that the cross structure is supe- 
rior to the circular and square shaped structures. Unfor- 
tunately, the length to width ratio in line structures is not 
very favourable for signalization. Given a minimum width, 
e.g. one pixel in image space, this implies relatively large 
targets. If the length to width ratio becomes smaller, the ad- 
vantage of crosses will disappear. In consequence, there is 
no reason to change the simple target shapes used so far. 
Only the size of the targets should increase to a certain ex- 
tend because this improves the measurement in the digital 
domain. 
For a further discussion on identification and discernibil- 
ity of simple shaped structures cf. Geiselmann and Hahn 
(1994). 
The templates used in the experiment are plotted in figure 
5, 
12: | 
t4 
t6: f 
toc: | 
  
Figure 5: The templates used in this experiment. 
Template t1 and t3 represent a square target and its un- 
sharp representation. By rotation of this templates t2 and t4 
are obtained. The templates t5 and t6 are discrete versions 
of circular shaped targets. The size of the bright region in 
this templates is about 5 x 5 pixels. Six further templates 
t1c to t6c are generated by calculating images of gradient 
strength of t1 to t6. 
In the experiment the templates are used pairwise, e.g. t1 
and t2, so that the matching with the minimum description 
length criteria described above leads to a decision on which 
of both templates gives the best measurement. 
294 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
3.2 Matching without Special Treatment of Background 
Noise 
If matching is carried out without giving respect to the back- 
ground problem the following result is obtained (table 1). 
Table 1: Matching result of template matching 
  
  
  
  
  
  
  
  
  
Template Matches 
type successful failed oo 
rel. [%] | abs. | abs. | [um] 
t1/t2 74.2 1272 442 2.79 
t3/t4 75.9 1301 413 2.85 
15/t6 55.5 952 762 2.90 
  
  
  
The measurement accuracy indicated by oo of the block 
adjustment is about 3 jm or 1/5 of the pixel size which is 
a quite good accuracy for signals with a size of 5 times 5 
pixels. For comparison a manual measurement is carried 
out for all 1714 image points and an accuracy of 3.17 um 
is obtained. The success rate of 75 % for the sharp and 
unsharp square is relatively low. And really bad is the 50 % 
rate obtained for the circle. 
3.3 Adjusting the Template Size 
To assess the dependency on the background the match- 
ing is carried out with various window sizes. Technically a 
circular weighting function is introduced into matching and 
binary weighting is used to eliminate all information outside 
a certain radius. 
Table 2: Dependency on the window size 
  
  
  
  
  
  
  
  
Templ. Matches 
type Rad. successful failed oo 
pixel | rel. [%] | abs. abs. | [um] 
t1/t2 5 99.4 1704 10 2.80 
6 99.5 1705 9 2.81 
7 99.1 1699 15 2.81 
8 98.7 1692 22 2.78 
9 98.7 1692 22 2.78 
10 98.2 1683 31 2.76 
  
  
  
  
  
  
  
  
The results (table 2) show at first a very high rate of suc- 
cessful measurements. The absolute number of failures 
is up to an aperture radius of 10 pixel only 31 (but jumps 
over hundred for radius 11). Thus for these signals the 
background of up to 4 or 5 pixels width around a target 
has no negative influence on the measurement of the sig- 
nal. An expected improvement of the overall accuracy ob- 
tained with increasing window size can not be observed 
from the table. But the obtained level of 2.8 um is 15 96 
better than the 3.2 um accuracy obtained for the manual 
measurements. The conclusion here is that the proposed 
procedure is able to measure the signalized points at least 
with the same accuracy as this could be done by a human 
operator at Digital Photogrammetric Workstations. 
  
   
  
   
      
   
     
    
    
    
     
  
     
   
   
    
   
  
  
  
  
  
  
  
  
     
    
   
    
   
   
   
    
    
    
   
    
    
The big 
of a blc 
togramr 
use con 
choosin 
called ( 
pressio 
depend 
25 
20 
d 
1C 
compression rate 1:X 
PI 
Figur: 
Of prin 
sion Ol 
rate. T 
plates 
successful matches [%] 
A det 
serve