Computer Vision-Based Intelligent Analysis System for Higher Education Classroom Behavior

This manuscript addresses the limitations of manual and subjective classroom behavior assessment in higher education by developing an intelligent behavior analysis model based on an improved pose estimation and temporal feature fusion network—high resolution network (HRNet). The enhanced HRNet architecture extracts fine-grained posture key points of teachers and students while a temporal convolution and self-attention fusion mechanism models key point sequences to capture dynamic behavioral transitions. A multi-task classification module is constructed to recognize multiple classroom behaviors in real time. Experimental results indicate that the improved HRNet achieves 88.2% key point detection accuracy under complex occlusions, with an average error of 6.3 pixels and a heatmap confidence score of 0.92. The behavior classification accuracy reaches 91.9%, and the F1-scores for listening and hand-raising behaviors are 0.95. The correlation coefficient between predicted classroom activity levels and manual annotations is 0.93, with an average error rate of 3.9%.