After having characterized each feature vector, we have to compare them together.
The best solution seems to be the Mahalanobis distance [14] involving
the covariance matrix of the vector components. It takes into account the way the
components change and the possible correlation
between each other. Moreover, this distance follows the distribution and a
thresholding can easily be performed in order to eliminate the worst matches. But the
results strongly depend on the estimation of this covariance matrix, which is quite
difficult to obtain with accuracy because of the high number of various samples
that its estimation requires. We will present in the next sections our solution to
compute the likeness of two vectors.