Automatic meeting analysis is an emerging research field. In this paper, we present stochastic algorithms for tracking people in multi-sensor meeting rooms, for a number of relevant tasks, including tracking multiple people, tracking head pose towards analysis of visual focus-of-attention, and tracking speaker activity using audio-visual information. A Bayesian framework based on Sequential Monte Carlo methods is used in all cases. We discuss the advantages and limitations of our approach, illustrate it with results, and highlight a number of open issues.