The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
¿.JL-'CSP 
4.3 Dense Grid Matching 
4.3.1 Ground Matching: The purpose of image matching is 
to extract object’s geometry information and confirm its space 
position. Space front intersection should be employed to 
calculate the space coordination of the corresponding points 
after obtaining parallax of left and right photo through image 
matching. The digital surface model is created after the ground 
coordinates of corresponding points have been obtained. Some 
inter methods have been employed in DSM generation, so the 
precision willreduce more or less. The image matching based 
on ground only needs calculating Z value because the planar 
coordinates (X, Y) are known. 
The vertical line locus (VLL) is adopted in the paper. If there is 
a vertical line locus on ground, its projection on the photo will 
also be a line (refer to fig 9). That is to say the projection of the 
intersection point A of VLL must be on the parallax line on the 
photo. To solute the Z value of point A on ground through 
making use of VLL to search the corresponding image points 
1 and 
Cl-s 
if 
Zi = Z min +i-AZ (/ = 1,2,-) 
.afX-X s ) + bfY-Y s ) + cfZ-Z s ) 
a i(X-X s ) + b 3 (Y-Y^+c^Zj -Z s ) 
.a 2 (y-y 5 ) + fe 2 (T-r 5 ) + c 2 (Z,-Z 5 ) 
afX-X^ + bfY-Y^ + cfZ'-Zs) 
.a\{X-X s ) + b[{Y-Y s ) + c\{Zi-Z s ) 
afX-X^ + b'fY-Y^ + c'fZ.-Zs) 
.a , 2 {X-X s ) + bI 2 {Y-Y s ) + c l 2 {Z i -Z s ) 
a’(X-X s ) + b' 3 (Y-Y s ) + c'fZ i -Z s ) 
(x„y,) => Pi 
p k > Pi then Z = Z k 
Figure 9, Sketch Map for Ground Matching 
The ground matching is a linear searching along Z with AZ in 
[Zmin, Zmax] because the planar coordinate (X, Y) is known. 
The search strategy can improve the space stereo matching 
success ratio. 
4.3.2 Multi-level Matching: According to the characters 
of the remote sensing image, multi-level matching strategy is 
adopted to overcome partial image distortion and facular 
infection and increase the matching speed. Image matching is a 
“sick” program, that is to say there would exist no 
corresponding point on one image for a given point on the other 
image because of occlusion or there are more than one 
corresponding points because of overlap or transparency of the 
object. In order to overcome the matter and make it an “order” 
program, additive condition and restrict should be imported. In 
order to control the complication degree and the reliability of 
the matching, it is very important to take into account the multi 
level data structure and geometry restrict. 
From low to high level strategy is adopted to generate remote 
sensing pyramid image and do image matching frequency 
divisionally. On the level 0 using single point grey matching 
which has a wider drag in range. The initial Z value of each 
point in the DEM is given the roughly average value of the area. 
Smooth is done to the DEM after each level matching is 
finished and the result is regarded as approximation value for 
the next level. The search rang includes all height distributing 
extent of the area. The higher level uses the results from the 
former level as its initial value and using shift spherical surface 
approach to forecast the result. The correlation coefficient is set 
a threshold and using statistical probability model to eliminate 
coarse error. 
Figure 10, The Image’s Pyramid Structure 
4.3.3 Matching window adjustment: It can refer to the 
reference [7] as matching windows self adjustment strategy is 
considered for high resolution stereo image matching. 
According to the multi-spectrum info character that remote 
sensing image has, the initial result which is classified from 
original image can be treated as the basis of window adjustment 
for the reconstruction of thematic element such as building. 
Low resolution multi-spectrum (RGB+NIR) image is used and 
spectrum character extraction is applied firstly, then max 
similar classification is adopted to realize image auto 
classifying to get original classification result [8]. The original 
classification result is used as covering template to distinguish 
foreground and background information. The image character 
information is integrated in the matching window to determine 
the image matching result and improve the matching reliability 
of character indistinctive area. 
(Left Image) (Right Image) 
Figure 11, The Spectrum Classify Results of the Stereo Images 
1369