AI-Assisted Activity Detection in K-6 Classroom Environments: A Preliminary Framework to Assist in Pedagogical Performance Evaluation

Video-based classroom observation tools can provide constructive feedback to K-6 teachers and assist their training and skill development. However, accurate classroom observation ratings necessitate observer training, requiring the investment of time, money and skilled labor. We propose a new AI-assisted pipeline to automate the classroom observation rating process utilizing a deep learning framework. Specifically, we combine three streams: action recognition, object detection, and age estimation networks to detect classroom instructional activities. We conducted the experiments on a novel labeled K-6 (elementary) classroom observation video dataset to detect human activities in the classroom environment. We labeled the data with the help of experienced annotators trained in classroom observation instruments. We also conducted multiple inter-rater reliability studies to ensure reliable labels. Our preliminary experimental results show promise for detecting multiple classroom instructional activity labels. © 2021 IEEE.