Multimodal perception and statistical modeling of classroom learning events: towards a privacy-preserving approach