re for 
roposed 
xterior 
hagery. 
values 
length. 
linear 
! Same 
data 
after 
multi 
it is 
esults 
h the 
e, the 
used. 
points 
ion of 
lermore 
‚es of 
ıt for 
lation 
) we 
ration 
80 be 
g the 
SPOT 
terior 
rate 
d of 
shading zones, in which the parallax can not 
be generated by matching algorithms or even by 
manual manner. So these areas in the image 
must be marked out before the matching 
processing, otherwise the reliable parallax 
grids can not be created. We have developed 
the quard- tree based seed region growing 
algorithm (Tao, 1992) to devide the image into 
different parts. The parts which belong to the 
dead-areas are marked with distinct signs and 
the parrallax values of the grid in these 
parts are interpolated by the neighbouring 
grid parallax values. 
Based on the level 1 processing results, the 
two-dimensional matching is used, whereas the 
searching window is much small, namely, 7 
pixels in line direction and 3 pixels in 
column direction. The most conjugate points 
are matched well in this sarching range. some 
points which are out of this range will be 
correctly determined by level 3 matching 
processing. In level 2 matching procedure. the 
multi-criterion matching method ( Lin, 1988) 
and smoothing technique based processing are 
applied for improving the matching reliability. 
It takes about 90 seconds to generate the 
parallax grids of image with size 512X512 in 
this procedure. 
3.3 level 3 image matching 
  
In order to extract the elevation information 
more precisely, the sub-pixel points matching 
algorithm must be applied. Least squares 
matching algorithm (Ackermann, 1983) was used 
widely in the field of photognmmatry since it 
can provide very high accuracy ( less than 0.1 
pixel). however, there are two weak points of 
LSM that cause a limitation in application: 
time consumption and small pull-in range . 
In fact, LSM can be treated as a adjustment 
system with additional parameters. From the 
additional parameters adjustment theory and our 
experiment results, we conclude that the main 
reasons of the above weakness are caused by the 
parameters selection in the LSM Generally, 
eight transformation parameters (six. for 
geometric and two for radiometric) for LSM are 
likely to be used. Theoretically speaking, the 
eight parameters are complete to compensate 
for the radiometric and geometric difference 
between the two correlation windows. In reality, 
however, in the case of small windows, the 
number of transformation parameters may be so 
sufficiently large that the correlation among 
the parameters occur to be put into effect. As 
a result, the model errors will be enhanced 
and the computation will become more time 
-consuming. 
875 
We have developed two modified LSM algorithm: 
SLSM and  PDFM. The former one is the 
simplified LSM model for which only the four of 
eight parameters are taken into use. The four 
parameters (ho, hi, dx, dy) which determine 
the parallax values directly can be called 
main parameters in this case. The other four 
parameters can be considered to be supplement 
parameters which compensate for the rotation 
transformation. SLSM may be written as: 
Vi = hoth;Ga (x, y! “Gm (x; y) (2) 
Ya = Ge (xt dx, y+dy) -Grs (x, y) (3) 
where (the  unknows can be successively 
approached by alternating the use of the above 
formula (2) and (3). 
For the purpose of improving the accuracy of 
the main parameters and reducing the model 
errors, we proposed a parameters dynamic 
filtering technique based LSM algorithm (PDFM). 
At first, the eight parameters are incorporated 
in adjustment while the conjugate points 
matching are active. After each time of the 
iteration , the parameters dynamic filtering is 
carried out by the t - hypothesis test. The 
statistical hypothesis is decsibed as 
followings : 
ti=a; / 0 - SORT (0 ;) 
g = SORT(X (AG)? / n-r ) 
A (i*G.- Ga. (4) 
where a, is the supplement parametar which has 
been used in the adjustment,and i denotes the 
ith supplement parameter. We take the signifi- 
cance level as the threshold for filtering the 
parameters. From the viewpoint of conservative 
-ness, &=10% is chosen and thus to ya is 1.65. 
After each iteration , once the t, is less than 
£a/2 ihe parameter a, is rejected from the 
succesive iteration of matching. 
The comparative experiment results are shown 
in Tab. 2 . SLSM has the advantages of the 
stability of accuracy, less consumption of 
running time and the large pull-in range as well 
(4-6 pixel). So SLSM is adopted to apply to 
level 3 matching procedure. In accordance with 
the criterion of speed, iteration times and 
accuracy, the PDFM is betler than LSM at all. 
The pull-in range of PDFM is enlarge also 
(about 2-3 pixel larger than 1~2 pixel of LSM). 
Therefore, the introduction of the parameters 
dynamic filtering technique is effective for 
improving the quality of LSM. 
43