Classroom Evaluation Based on Behavior Recognition and Stacked Hybrid Experts

With the continuous development of intelligent education, achieving efficient teaching quality evaluation in large-scale classrooms has become an important research direction. To address the limitations of traditional approaches that over-rely on expert judgment and suffer from low efficiency, this paper proposes a classroom score prediction method based on behavior recognition and a stacked mixture-of-experts model. Built on an improved YOLOv10n-LD, the system recognizes four student behaviors in classroom videos—Listening, Bowing, Sleeping, and Using Phone—and computes behavior occurrence rates from frames sampled every 30 s. From these rates, twelve statistical features (such as mean, variance, extremes, and skewness) are extracted and fused with Class size to form a 49-dimensional feature vector. A feature selection mechanism finally retains fourteen optimal features, including Class size and the Listening rate, which are fed into the Stacked-MoE for classroom score prediction. Experiments show that the method achieves RMSE, MAE, MSE, and MAPE of 2.6028, 2.009, 6.7744, and 2.5513%, respectively, outperforming traditional baselines. This study provides an effective solution for intelligent evaluation of university classroom teaching quality and shows strong application prospects.