Boosting and Structure Learning in Dynamic Bayesian Networks for Audio-Visual Speaker Detection