CNNs and Transfer Learning for Lecture Venue Occupancy and Student Attention Monitoring

Lower student success rates in higher education might, in some case, be due to the unprecedented increase of student numbers without a comparable increase in resources and funding. This paper proposes a face-based detection system to monitor occupancy and student attention in crowded classroom using state of the art deep Convolutional Neural Networks (CNN) architectures. The aim of the proposed system is to contribute to the increase of subject success rates by monitoring attendance and attention. The system utilizes a two-phased approach: The first phase determines the number of student faces in an image frame. The Haar Cascade, LBP, HOG, Resnet CNN, TinyFace CNN, and SSD were compared to determine the algorithm best suited to the detection of faces in crowded classroom scenes. In phase two, the orientations of the faces are determined using transfer learning. Faces are classified as “right”, “left”, or at the “center”. This information is displayed on an augmented reality display to provide feedback to lecturers in semi real-time. It is hoped that this will assist lecturers to address problems related to student attention in crowded classrooms.