ISPRS Workshop on Service and Application of Spatial Data Infrastructure, XXXVI(4/W6), Oct. 14-16, Hangzhou, China 
the two images. It is the basic idea of LSM to estimate the 
parameters of transformation from these observed gray level 
differences by a least squares adjustment. The mathematical 
model can be expanded to other transformations and to handle 
also radiometric parameters. Due to the great number of 
observations, LSM is the most accurate image matching 
technique. However, it is very sensitive with respect to the 
quality of the approximations. 
In this paper, we are not contemplating the development of any 
new subpixel accuracy techniques. However, as getting such 
additional resolution effects the execution time of time of image 
matching, we will use correlation interpolation in 1D matching 
and Least Squares matching in 2D matching to achieve subpixel 
accuracy. 
2.8 2D matching 
For reduce the search from 2D to ID, we need computation of 
the unknown parameters of the epipolar geometry that needs a 
certain number of well-distributed conjugate points. These 
conjugate points are extracted by the following 2D matching: 
The SUSAN interest operator is used to select well-defined 
feature points that are suitable for image matching; Conjugate 
points are generated using modified hill climbing method 
(Zhang, 2005). The positioning of the search areas is 
determined by using the already known control points. For 
reliability, the threshold of acceptable normalized correlation 
coefficients is 0.9; Delete bad match point using epipolar 
geometry; Least squares matching is finally used to refine the 
image coordinates of these points in order to achieve subpixel 
accuracy. This procedure results in several hundreds of 
conjugate points. 
The distribution of the conjugate points is the figure 5. 
2.9 Quasi-Epipolar Image Generation 
Unlike frame-based imagery, where all pixels in the image are 
exposed simultaneously, each scan line of the IKONOS image 
is collected in a pushbroom fashion at different instant of time. 
Thus, epipolar lines with linear CCDs become curves. A 
straight line for a small length but not for the entire image can 
approximate it. An epipolar curve for the entire image can be 
approximated only by piecewise linear segments. In our 
approach, the epipolar curve(Jiang,2002) for point ^ 1 ^ 
in the left IKONOS image is approximated by a quadratic 
polynomial and has the following form: 
y r =a 0 + a x x r + a 1 y l + a 3 x r x r + 
a A x r y, + a s x r x r y l 4- a 6 x r x r x r (l l) 
+ a 7 x r x r x r y, 
Where the ^ r ^ are the pixel coordinates in the right 
Cl rv Cl n 
IKONOS image and u are unknown parameters. Using 
the epipolar geometry, the quasi-epipolar image pair can be 
generated by re-arranging of the original IKONOS image pair. 
However, the computation of the unknown parameters of the 
epipolar geometry needs a certain number of well-distributed 
conjugate points. These conjugate points are extracted by the 
2D matching. 
2.10 Fast Image Correlation 
The most time-consuming step in image matching is calculation 
normalized cross-correlation, but the calculation has a lot of 
redundant computation. To relieve this redundancy, the 
necessary storage space is retained for the calculated 
intermediate results. These stored results can be used directly. 
Effectiveness will increase four to ten times according to the 
size of the match window. 
2.11 Matching Procedure 
The match processing is an iterative procedure, and can be 
formulated as follows: 
• 2D matching procedure results in several hundreds of 
conjugate points; these points can be used to recover the 
epipolar geometry and interpolate the approximate values 
for the following point matching procedure. 
• Quasi-Epipolar Image Generation 
• The match points are selected and distributed in the form 
of a regular grid in the left image or interest points which 
extracted by SUSAN. Given a point in the left image, a 
search window in the right quasi-epipolar image can be 
determined by the 2D matched points. The correct match 
of this point should lie in this search window. However, 
due to repetitive texture or poor texture information, there 
could be several candidate matches appearing in the 
search window. These candidate matches are located 
along the epipolar curve. They can be derived by 
normalized cross-correlation technique, and the candidate 
matches are selected if their correlation coefficient lies 
above a certain user-defined threshold. We use the fast 
image correlation in this step. 
• Achievement subpixel accuracy of all candidate by 
correlation interpolation 
• Compute the matching strength for each candidate match 
• Update the matches by modified greedy algorithm 
• Check the false match by the local reliability constrain 
• Interpolate the null match point by finite element 
The parallax map of the IKONOS images by our image 
matching is the figure 6. 
3. TEST DATA 
IKONOS images over Dutch were used for the experiments. 
The left and right pair was acquired on the same day of. The 
size of each image was 6560by 7556pixels. The images cover 
approximately 8kms by 8kms on the ground. Figure 1 shows the 
IKONOS image (left). The image contains dense population of 
residential houses, apartments and industrial buildings as well 
as rivers and hills. Automated matching such images is a 
challenge to any stereo matching algorithms. 
4. CONCLUSION 
We have proposed in this paper a new approach to match two 
images by both epipolar constraint and local reliability 
constraint with modified greedy algorithm and new support of a 
match. The method has been experimented on some real