Classification de signaux multidimensionnels utilisant la distribution de Wishart : Application à la reconnaissance de mouvements

Overview & Synopsis

This paper proposes a novel use of the Wishart distribution within the scope of supervised or unsupervised movement recognition and clustering. As a movement signature, we retain the covariance matrix between spatial coordinates of sensors, conveniently fastened on the performer joints. In both cases, each movement type is supposed to be associated with a special Wishart distribution, whose parameters are liable to be estimated with maximum likelihood technique. The assets of these two approaches are on the one hand (supervised), to provide satisfying rates of recognition, and on the other hand (unsupervised) to point out a finer taxonomy of movement. We assert the robustness of the method when clustering the data with several smoothing levels. We eventually present with the results obtained on a set of 219 classical dancing movements