The images are locally normalized channel by channel in order to obtain the invariance to change of intensity
through the model described at equation 4. The width of the window depends on the
differences between the images contents.
The derivatives are computed in a stable way till order one with a Gaussian filter and a sub-pixel precision
[11] on
these normalized images. The matching of the
feature vectors of the equation 3 obtained is done by using the
techniques introduced at section 4. The points to be matched are the Harris points
presented at section 3.