Spatio-temporal object recognition using variational learning of an infinite statistical model

In this paper we present a sophisticated variational Bayes framework for learning infinite Beta-Liouville mixture models. A key feature of the proposed framework is that the appropriate mixture model complexity can be discovered automatically from the data to cluster as part of the inference procedure. Another important advantage is that the whole inference process itself is analytically tractable with closed-form solutions. Moreover, the problems of over-fitting and under-fitting are also prevented thanks to the nonparametric Bayesian nature of the proposed framework. The effectiveness of our statistical framework is investigated on two challenging motion recognition tasks including hand gesture and human activity recognition.