Speaker tracking based on audio-video information fusion