3.2 Character Points Matching 
Making use of the grey information round the acute point and 
the correlation method to build a local matching rule so as to 
divide the results detected by Harris operator into many-to- 
many matching pairs. The correlation coefficient is adopted to 
describe the measurement of similar degree of point P on 
image A and point ^ on image B. The correlation coefficient is 
defined as [2]: 
Correip, = Yj-Au+Î,V+j)- /4 ]■ [B(u+i,v+j)-^\ (3 ) 
C>]C>2 j=-ni=-n 
o 
( P" 1 ) and 1 ( 0,2 ) express the mean and the square 
difference of point p (q) on image A (B) separately. The n is 
the neighbouring range of the acute point. 
Figure 4, The matching results of the characters on stereo image 
pairs 
An initial set of 'corresponding points is gained through 
character points matching, but if only grey character is relied as 
the correlation measurement there will exist some error 
matching points(refer to Figure 4, the matching results are not 
exactly corresponding), so that eliminating coarse error should 
be evolved. 
3.3 Epipolar Detection 
The corresponding points on the stereo image pairs should 
satisfy the coplanar condition according to the geometry 
relation of the imaging [3]. 
B B 
x j 
7\ 
J 2 7: 
defining: 
z t x 2 - x x z 2 
Is, aJ 
\X t Z,] 
i k -k m 
— Y. - 
\x 2 Z z j 
z,x 2 - x\z 
i-0, B 
W-1 z. 
z,x 2 -x x z 2 
= -N t i\ + NX 2 
= -0 
Q is the model fluctuation parallax of the orientation points. 
When a stereo image pair have been done relative orientation, 
the value of the Q is 0, otherwise, Q^O. That is to say the 
corresponding points must fall on the corresponding epipolar 
lines. This rule is used to eliminate the wrong matching point 
pairs in the initial corresponding points set. 
The basic principle of the least squares image matching can be 
described as following [2]: the grey and geometry aberration 
parameter are imported when doing image matching, and these 
parameters are calculated through least squares image matching 
principle to improve the precision of image matching. 
The simple aberrance is considered in practice because the size 
of the matching windows is often small. 
x 2 = a 0 + a t x+a 2 y 
y 2 = b 0 + b t x + b 2 y 
If the linear grey aberrance of the right image relative to the 
left one are taken into account, then the following formula can 
be acquired: 
g, O, y) + «, (x, y) = h 0 + h x g 2 (a 0 + a,x + a 2 y, b 0 + + b 2 y) + n 2 (x, y) 
(7) 
The error formula of least squares image matching can be 
acquired by establishing error formula and do linearization to 
the formula. 
v = c x d\ + c 2 d\ + c 2 da 0 + c 4 da x + c^dc^ + c 6 db 0 + c n db\ + c 8 i/6, - Ag 
(8) 
Establishing error formula pixel by pixel in the object area and 
solving the grey and geometry aberrance by normalizing. Then 
geometry and grey transformation is applied to right window 
to get new image, calculating iteratively the correlation 
coefficients of the left and the right window, the object place 
is obtained when the value of correlation coefficients do not 
rise and the loop course to be stopped. During the loop course, 
the threshold is set as a condition to eliminate coarse error and 
eliminate the unreliable matching points.