Barnea and Silverman [5] introduced a class of sequential algorithms for fast image registration. They were designed to reduce computation in matching procedures using minimum dissimilarity measures like the sum of the absolute differences (SAD). Konecny and Pape [6] reviewed image correlation techniques according to photogrammetric and mathematical fundamentals.
Different similarity measures have been used in the literature [7, 8], and their performance and computation cost vary. It has also been shown that the zero mean normalized cross correlation and the zero mean sum of squared differences tend to give better results [9]. We will use the zero mean normalized cross-correlation (ZNCC) coefficient as the similarity measure of the candidate matching areas. The estimate is independent of differences in brightness and contrast due to the normalization with respect to mean and standard deviation.
Let 
be the intensity value of an 
image f at
position (m,n), where f is to be box filtered into 
,
i.e. obtaining the mean of the original image within the box. We also
have similar definition for a second image g. The
normalized cross-correlation of two windows can be written as follows:
where
and d is the shift along epipolar lines; K and L define the correlation window size. It can be seen from this equation that the co-variance between f and g and the variances of f and g at different positions in the image need to be evaluated.
