Building Interpretable Descriptors for Student Posture Analysis in a Physical Classroom

Abstract

This research presents a process for simplifying video labeling and feature generation when building classification systems from real classrooms. Using video from a single, wide-angle recording of a live classroom, we create a lowlevel feature set of posture primitives built on keypoints from OpenPose. We use that feature set to build a posture recognition model of “natural labels” built from a scripted posture video using the same classroom. This model provides automatic labels for the real classroom data. We then derive a set of interpretable descriptors to characterize student-specific posture pattern dynamics. We show that those descriptors are able to discriminate between subtle differences in learning activities in a real college classroom.