Audio-visual tracking of a variable number of speakers with a random finite set approach.