Multimodal perception and statistical modeling of classroom teaching events: toward a privacy-preserving approach