A new dataset and systematic evaluation of deep learning models for student activity recognition from classroom videos

Student activity, a non-verbal modality, is an important cue to imply the students’ learning state and engagement level during class. Quantitative observation of the student activities will contribute to innovating teaching methods. However, traditional methods that rely on manual observation are time-consuming. Nowadays, surveillance cameras have been widely deployed in the classroom. This results in a massive amount of classroom videos. However, the exploitation of these videos is limited due to the lack of annotated information. This paper has two main contributions. First, a fully annotated dataset dedicated to student activity analysis has been proposed. To this end, five activities in the classroom context including sitting, standing, sleeping, raising-hand and using-phone have been defined and collected in different class sessions. Second, three deep learning models have been fully assessed for the activities recognition.