Predicting classroom activity index through multi-scale head posture classification network.

The quantification of the interplay between student behavior data and classroom teaching effectiveness using quantitative metrics has perennially posed a challenge in the evaluation of classroom instruction. Classroom activity serves as a reflection of student engagement, emotional ambiance, and other pertinent aspects during the pedagogical process. This article presents a methodology for quantifying student head posture during classroom instruction utilizing AI-driven video analysis technology, notably the Classroom Activity Index (CAI). A Classroom Activity Analysis System (CAAS) was designed and developed, integrating a multi-scale classification network based on ECA-ResNet50 and ECA-ResNet18. This network discerns and categorizes various head regions of students situated in both the frontal and real rows of a lecture-style classroom, irrespective of their dimensions. The classification network attains exceptional performance, boasting F1 score of 0.91 and 0.92 for student head-up and head-nodding. Drawing on the live classroom instruction at a higher vocational college in Wuhan, Hubei Province, China, a comparative experiment was executed. The findings revealed that three factors: teacher-student verbal interaction, teacher body language, and utilization of digital resource, all exert an influence on CAI. Simultaneously, the degree of classroom activity as gauged by FIAS and manual analysis fundamentally aligns with the CAI indicators quantified by CAAS, validating the efficacy of CAI in the quantification of classroom activity. Consequently, the incorporation of CAAS in teaching, research, and oversight scenarios can augment the precision and scientific rigor of classroom teaching assessment. [ABSTRACT FROM AUTHOR]