Toward Automated Classroom Observation: Classification of Elementary Instructional Activities Using Neural Networks

• In Event: Virtual Agents, Artificial Intelligence, and Automation

Abstract

Classroom videos have become a hallmark for teacher education and professional development. These videos are also critical to appraisals of teaching. A limitation of using classroom videos for assessing teaching quality on a large scale is the financial and time-consuming costs of amassing human raters. Recent advances with computer vision-based human activity recognition (HAR) systems and neural networks, however, may offset some of these costs. However, these technologies need large video datasets with reliable and valid activity labels for training. In this paper, we describe a 263-hour dataset of classroom videos with 24 instructional activities applied to mathematics and English language arts (ELA) instruction in elementary classrooms and the outcomes of a neural network developed to identify these instructional activities.

Authors

• Jonathan K. Foster, University of Virginia
  Presenting Author

• Matthew Korban, University of Virginia
  Presenting Author

• Peter A. Youngs, University of Virginia
  Presenting Author

• Ginger S. Watson, Old Dominion University
  Presenting Author

• Scott T. Acton, University of Virginia
  Presenting Author