Comparison of video dynamic contents without feature matching by using rank-tests
Résumé
This report presents a novel and efficient dissimilarity measure between video segments. We consider local spatio-temporal descriptors. They are considered to be realizations of an unknown, but class-specific distribution. The similarity of two video segments is calculated by evaluating an appropriate statistic issued from a rank test. It does not require any matching of the local features between the two considered video segments, and can deal with a different number of computed local features in the two segments. Furthermore, our measure is self-normalized which allows for simple cue integration, and even on-line adapted class-dependent combination of the different descriptors. Satisfactory results have been obtained on real video sequences for two motion event recognition problems.