Online Analysis of Hierarchical Events in Meetings

摘要

Automatic online analysis of meetings is very important from three points of view: serving as an important archive of a meeting, understanding human interaction processes, and providing the attentive services based on the meeting situation for participants. Based on this view, this paper presents principle and implementation of online analysis of hierarchical events in meeting scenario. A hierarchical dynamic Bayesian network modeling different levels of events is designed. In this model, the recognition of low-level events is supervised by high-level events Rao-Blackwellized particle filter is proposed for on-line inference for the hierarchical dynamic Bayesian network. Situation events and four sorts of interaction events in meeting scenario are detected and recognized. Experimental results show that our approach can detect and recognize multi-layer semantic events in dynamic environment. Comparing with previous methods of meeting analysis, our approach supports online probabilistic inference for activities at different layers in meeting scenario.