The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
X r =[Xlxlr-X\I; /=[/„/ 2 ,A= 
a i ••• 0 
o - 4, 
(8) 
x=XX-XT; XLL’Xj; b 
3 ••• o 
So the gray observation equation (1) can be expressed as 
v(x, y) = AX—l\P (the weight coefficient matrix ofl) 
(9) 
3.2 The Geometrical Observation Equations 
The qusi-epipolar line is used as constraint factor and maked up 
of the geometrical observation equations in this paper. As is 
shown in Figure 1, the corresponding point 
Pi,i = 1,2> 2) in searching image must locate at 
their corresponding qusi-epipolar line, therefore these qusi- 
epipolar lines can be used as geometrical constraints and form 
corresponding geometrical observation equations to join the 
adjustment.Generally, the qusi-epipolar line can be expressed 
by the formation of polynomials 
yi = /(*,■) = a im x? +.: + a n x i + a i0 (10) 
For the L0 level images of the linear array sensor, the qusi- 
epipolar lines are not the straight lines generally. But for the LI 
level images, the qusi-epipolar line can be simulated by a 
straight line. Then Equation (10) becomes 
yi = f(x,) = a n x i + a i0 (11) 
If we have get the approximate pixel coordinate ( X ; °, y i ) of 
the image point by AMMGC, the geometrical observation 
equations can be expressed by 
V(x,y) = Ay, - a n Ax, + O, 0 - /(x, 0 )) (12) 
With the notations 
XJ = [Ax ; , AyJ 
•i ei =-y, 0 +f(A) 0 = 1,2,..«) 
»= [”«„,!] 
Here (Ax ; ., A_y ; . ) is corresponding to (Aw ( , AV ; .) gotten in 
the gray observation equations, so the geometrical observation 
equations can be given by 
v(x,y) = BX ~lg\Pg (the weight coefficient matrix of / g ) 
(14) 
Gray observation equations (9) and geometrical observation 
equations (14) compose a combined adjustment system, which 
are associated with each other via the common conversion 
parameter (u j ,V ( .) .The least square solution of combined 
adjustment is 
X = (A 7 PA + B T P g By\A T PI + B T PJ g ) (15) 
The answer of equation (15) is the final matching points 
coordinates, which can reach the 1/10 pixel matching accuracy 
in theory. 
4. DSM GENERATION 
Based on image pyramids and feature points, DSM are 
automated generated in this paper. Usually, feature points 
correspond to the points with acutely changing intensities, so 
they are fairly important for DSM generation. Additionally, 
feature points are precise and reliable, and they can be extracted 
by many methods. 
However, in the image areas with coarse image textures and 
even no image textures, feature points can not be extracted. 
Therefore, apart from feature points, grid point has to be used to 
ensure the generation of precise and dense DSM. Grid points 
are the points evenly distributed among images. Compared with 
feature points, grid points may lie at image areas with coarse 
texture or even occluded and thus will obtain incorrect 
matching results. 
On each image pyramid layer, matching results for feature 
points and grid points are obtained through multiple image 
matching algorithm. Feature points are matched through 
approximate DSM obtained from higher image pyramid layers 
and the matching results are used to constraint the searching 
distance of the grid points. In order to ensure the quality of the 
DSM, dense feature points and grid points have to be obtained 
in the matching process and the process is that as follows. 
Firstly, feature points with high interest values are extracted by 
Fòerstne operator. Then, these points are matched by traditional 
2D searching methods. Successful matching results are used to 
interpolate y-parallax gird, which are used to compensate the y- 
339