some mismatching results caused by the noise or salient 
points still exist, we can use this characteristic to 
confirm the real linear feature and get rid of the false 
matching pairs. 
The result of string matching was found to be highly 
reliable. The accuracy is not as high as intensity-based 
matching, but it can offer good enough pull-in range for 
later refinement by object space least squares matching. 
5. RESULTS OF EXPERIMENT 
(1) For epipolar lines resampling, relative orientation is 
performed with Kern DSR-1 and rotation elements of 
left/right images are obtained. After the epipolar lines 
resampling, the conditional rankorder filter is applied to 
remove minor features and part of noise (Fig. 4). 
NS 
NS 
  
Fig. 4: The left image after epipolar lines resampling 
and conditional rankorder filter 
(2) For linear features detection, the gradient filter is 
applied and linear features are located by zero crossing. 
(Fig. 5) 
(3) By string matching, the false linear features which 
show up in one image only are removed, and the 
conjugated linear features which show up in both 
left/right images are extracted. (Fig. 6). According to 
these useful conjugated linear features pairs, the coarse 
DEM data can be produced. 
476 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
  
Fig. 5: The raw linear features detected by gradient 
filter (left image) 
  
Fig. 6: The useful linear features extracted by string 
matching (left image) 
   
After 
the lin 
is app 
inform 
used : 
these 
range, 
perfor 
start ¢ 
lead t 
would 
fine E 
can be 
The | 
match 
middl 
match 
deterr 
soluti 
featur 
decisi 
decisi 
a cor 
match 
for D 
(a) F 
level 
not 
radioi 
rather 
match 
pairs 
appro 
(b) F 
area ( 
matck 
result 
from 
matck 
but al 
cost € 
(c) TI 
by us. 
and « 
meas 
primi 
primi