vhich 
them 
are 
The 
ne 
ed to 
ne 
he 
ixels 
ters 
sresti 
input 
jints 
the 
2.7.2 Defining Initial Matches 
The first step is to determine possible matches 
between an interest point in the first image and 
any number of interest points in the second 
image. A geometric transformation exists 
between the two images, supplied by a higher 
level in the pyramid or by the operator. This is 
used to predict the position of a point from the 
first image, in the second image. From this 
predicted position a search is made within a 
window of specified size for interest points. 
Any points found inside this window are 
considered as possible matches. 
The result of this procedure is a list with the 
identity of a point from the first image 
associated to the identity of possible matches in 
the second image. It should be noted that a 
single point in the first image can be matched to 
several points in the second image, and vica 
versa. 
2.7.3 Cross Correlation 
The above procedure tends to produce a large 
number of matches, some of which are 
ambiguous. Considering a given conjugate point 
pair, an image chip is placed around each pixel 
and image cross correlation is calculated. This 
may require the resampling of one of the chips 
if the is significant rotational or scale 
differences between the two images. If the 
cross correlation value falls below a given 
threshold then this conjugate pair is removed 
from further consideration. 
131 
2.7.4 Robust Transformation 
Estimation 
The resultant conjugate points from the above 
procedure are then processed by a Robust 
Transformation Estimation procedure. This 
procedure seeks to determine a geometric 
transformation between the two images based 
on the conjugate points. It is an iterative 
solution in which each conjugate pair is initially 
weighted. These weightings are updated based 
on the calculated point residuals for each 
iteration. Outliers are removed at each 
iteration, and the procedure continues till 
convergence is achieved, or a maximum number 
of iterations has occurred. The initial weights 
of the conjugate points are based on each points 
‘weight’ calculated by the interest operator. 
The robust transformation estimation results in 
a trimmed set of conjugate points, in which 
unlikely conjugate points having been rejected. 
A new geometric transformation is also 
defined. 
It is likely that some ambiguities still exist, 
with a single point in the first image matched to 
more than one point in the second image, or 
visa versa. These ambiguities are removed by 
selecting the conjugate point pair with the 
lowest residual (weight). Once ambiguities 
have been removed a new geometric 
transformation is defined. 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996