Because our system is designed as a real-time application
we cannot try different sizes of the detected object (patch) in
order to verify which classifier offers the best result and then
make a decision. Moreover, when resizing the patch to a
specific dimension in most of the cases the image is distorted.
So, in order to avoid these problems we propose not
constraining the patch size but to limit the total number of
cells for which we compute the HOG descriptor. In this case,
the image is resized but the aspect ratio is preserved and the
descriptor needs to be extracted only once.