Extraction of mid-level semantics from gesture videos using a Bayesian network