Evidence has shown that student's attention is a crucial factor for engagement and learning gain. Although it can be accurately assessed ad-hoc by an experienced teacher, continuous contact with all students in a large class is difficult to maintain and requires training for novice practitioners. We continue our previous work on investigating unobtrusive measures of body-language in order to predict student's attention during the class, and provide teachers with a support system to help them to "scale-up" to a large class. Our work here is focused on head-motion, by which we aim to mimic large-scale gaze tracking. By using new computer vision techniques we are able to extract head poses of all students in the video-stream from the class. After defining several measures about head motion, we checked their significance and attempted to demonstrate their value by fitting a mixture model and training support vector machines (SVM) classifiers. We show that drops in attention are reflected in a decreased intensity of head movement. We were also able to reach 65.72% correct classifications of student attention on a 3-point scale.