Real-time situation detection based on Rao-Blackwellized particle filters in meetings

Real-time detection and recognition of situation events is very important for meeting analysis and archive in dynamic meeting environment. A meeting is modeled as a sequence of meeting situation events, which denote meeting states. Four high-level situation events, that is, "monologue", "presentation", "discussion" and "break", are considered and analyzed in dynamic meeting scenarios in this paper. A hierarchical dynamic Bayesian network is constructed to bridge the gap between low-level physical audio-visual features and high-level semantic concepts. Rao-Blackwellized particle filter (RBPF) is proposed for on-line inference over the state spaces of the hierarchical dynamic Bayesian network In order to efficiently evaluate the performance of our RBPF algorithm, comparative experiments are made using the regular particle filter (PF) method, in addition to the RBPF method. Experimental results show that the RBPF method achieves better the performance of real-time event recognition than the PF method.