Unobtrusive Classification of Student Behavior in Real Classroom Settings Using a Video Dataset

The recognition of human actions in videos is an important area of focus within the realm of artificial intelligence (AI) research. While benchmark datasets for activities such as sports and daily routines have shown remarkable improvement in various areas, there remains a noticeable lack of video datasets in the educational domain, particularly concerning student behavior in real classrooms. To address this gap, this paper introduces the Students Class Behavior Video Dataset (SCBVD), contributing to the expanding domain of human action recognition datasets tailored for educational contexts. The SCBVD dataset empha-sizes actions encompassing student behaviors in real classroom settings. Organized into ten distinct action classes, it enables a holistic and focused study of student behavior within the education field. To ensure diversity, each action type comprises no less than 200 clips, resulting in a cumulative duration of approximately 6 hours. Leveraging the Long Term Recurrent Convolutional Network (LRCN) model implemented with the Keras tuner, our evaluation on the SCBVD dataset yielded a classification accuracy of 74.04%. We further benchmarked this against popular video benchmark datasets such as UCF50 and JHMDB using the same model. Our findings underscore the potential of the SCBVD dataset in advancing research on student engagement monitoring systems.