The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
1368 
(Left Image) (Right Image) 
Figure 6, Superior Results by Using Least Squares Image 
Matching(The Orange Points are the Original Position of the 
Character, The Red Points are the Position of the Characters 
be Disposed Superiorly) 
3.5 Iterative Stereo Orientation 
XV XV 
The corresponding coordinates ( 1 , • 71 )and( 2 , 2 )are 
obtained by accurate measurement, then resolving the relative 
B X ,B V ,B Z , A a x , A co, Ak 
orientation elements ( y ), the 
model coordinates are obtained by front intersection formula 
and then the stereo model is reconstructed. 
During the program from character point detection to 
orientation, a series of criterion is adopted as conditions to 
insure the correct of process. The criterion includes point 
character value, correlation coefficient and fluctuate parallax. 
Because the process following is a sequence one, the criterion 
used in each phase is localization which results in the quantity 
and distributing of the results can’t satisfy the requirement of 
high precise orientation. The iterative strategy is adopted to 
settle the localization problem, which repeat step two to four on 
basis of the initial orientation. If integrated criterion is used in 
processing the effect of the stereo image auto-orientation can be 
improved farther. 
4. DENSE IMAGE MATCHING TO AUTO-EXTRACT 
DSM 
“linearity factorial” . t ^ e t j me complication degree of 
the web construction algorithm is “power factorial” 
0(fl\ogYl) . corn pij ca tj 0n degree of search time of the 
triangle cell or web vertex is “logarithm factorial” 
which is the most optimized resolution for such kind of 
question as query and search. The application effect refers to 
Figure 7. 
Figure 7, The Irregular Triangular Web Constructed by 
Corresponding Character Points 
4.2 Iterative Edge Character Matching under Space 
Restriction Condition 
The initial irregular triangular web obtained by image character 
points matching is rather coarse because the infection of 
adverse factors such as occlusion, noise, shadow and aberrance 
etc. The space restriction condition needs to be extended and 
perfected for dense image matching. In order to avoid resulting 
in wrong results, the edge character matching is adopted to 
obtain typical space character lines. The lines are used to 
restrict the error in the DSM and improve the accuracy of DSM. 
Digital surface model (DSM) is extracted by dense image 
matching. The initial irregular triangular web constituted with 
corresponding points gotten from the stereo orientation is used 
as geometry restriction conditions. The self-adaptive window is 
used in the matching course to settle the edge fuzzy in the high 
resolution image matching. The selection of self-adaptive 
window makes use of the classify results of multi-spectrum 
images. The image in the matching window is divided into 
background and thematic information to eliminate the effects of 
the discrete ground surface. The final digital surface model is 
the refined of the spectrum classify and the thematic 
information levels such as building and vegetation are 
generated. The basic technique following refers to the down 
part of Figure 3. 
4.1 Irregular Triangular Web Construction by 
Corresponding Points 
An increment style random dynamic Delaunay triangular web 
construction algorithm is promoted in the paper according to 
how to make use of discrete corresponding image points to 
quickly construct Delaunay triangular web with optimize space 
structure and how to do effective edition (such as add or delete 
vertex) on the already exist web [4]. The algorithm has such 
characters: adaptability, randomicity and high efficiency 
dynamic edition for great capacity of data. The capability of the 
algorithm is analyzed thoroughly: the space complication 
degree of the data structure involved in the algorithm is 
There are abundant research on edge extraction and matching. 
The Canny operator is adopted in the paper to detect edges [5], 
tracking the edges to extract the vector character of the edge 
after burr having been wiped off. The inflexion of the edge 
character can be confirmed when the grad value of the point is 
greater than a threshold. After inflexion obtained, the inflexions 
on the character lines can be matched with the epipolar 
condition to obtain corresponding points. Creating web with 
inflexions obtained to establish triangular web. Based on the 
relative connection of these grids to restrict the search range in 
line character matching and make edge matching quick and 
exact. Character matching methods refer to reference [6]. 
Figure 8, Vector Edges Extraction from Stereo Image Pairs