1 as false or 
ise compo- 
tion on the 
ita acquired 
ne, i.e. with 
polar lines 
chle, 1990) 
e short and 
ability and 
currence of 
spondences 
does not try 
ring ambi- 
22a) that if 
length ratio 
achieving a 
matches is 
is also been 
stereo by 
off between 
interesting 
, which is 
d matching 
inge photo- 
vipolar line 
> derivation 
iages of the 
. 8090/6096 
aken from a 
] stillvideo 
Swiss Alps, 
400m, were 
50mm lens. 
0% and the 
large devia- 
k-and-white 
(42 micron 
ixel) on an 
Ihe fiducial 
with least- 
rmation for 
the limited 
ne intersec- 
    
Ge Hm" m HW ANE AmECCHNE ANC ANE UNE 
Ce w- w-— w- - w cH e m À 
Figure 2: Simplon area scanned aerial image, common region marked 
tion technique a sub-block of 2 x 3 images was taken out 
of the project, with approximately 36% of the image 
format covering a common region. The camera orientation 
data was taken from a former aerial triangulation of the 
whole block on an analytical plotter. 
Discrete points were extracted from the six images of the 
sub-block by an interest-operator (Foerstner, 1986). 
  
Figure 3: Subregion with interest-operator extracted points 
With the Foerstner-operator between 50'000 and 80'000 
image points were extracted in the overlapping regions of 
the six images, which were slightly enlarged to 45% of the 
whole image area to allow for slight deviations from the 
flight plan and for effects of the topography on the size of 
the image overlap. These points were fed into the epipolar 
line intersection routine, which was able to match a total 
of about 1500 points in all images. The original result 
without any post-processing is shown in Figure 3. Note 
that no post-processing has been applied to these data. 
With only about 1500 points matched in all 6 images the 
yield is relatively sparse and not quite satisfactory. On the 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
other hand, however, the blunder rate is extremely low as 
compared to other automatic DEM generation techniques: 
In Figure 4 only 5-6 blunders can be detected, which 
corresponds to a blunder rate of only 0.3%. Moreover, the 
blunders show up as clear peaks or holes in the visualiza- 
tion can easily be removed by local post-processing 
methods like median filtering or robust surface fitting. 
roo Ase Mn RT 
  
uS 
Figure 4: Simplon test area, 10m 
contour lines (orientation 
and scale do not match 
with Figure 2) 
example of 
a blunder 
  
When in addition to the points which were detected in all 
six images also those points were accepted which were 
matched in any 5 of the 6 images, 5500 points could be 
matched, but with a larger percentage of blunders of 
approximately 2%. The relatively sparse result can be 
explained by the following two reasons: One problem in 
the data processing was the poor geometric stability of the 
uncalibrated scanner, which necessitated to work with a 
rather large tolerance to the epipolar lines (80 micron in 
this case) to make sure that corresponding points were not 
rejected due to inaccuracies in image space. As the 
number of ambiguities in the establishment of correspon- 
dences grows with the square of the tolerance to the 
epipolar line, this fact leads to an increasing number of 
correct matches being rejected as spurious matches due to 
unsolvable ambiguities. Another reason is the probability 
of the interest operator detecting identical points in all 
involved images, which decreases exponentially with the 
number of images. 
487