VMAPN: A Vision Mamba-Based Real-Time Behavior Recognition and Feedback System for Smart Education

Abstract

With the growing demand for intelligent educational technologies, traditional methods of classroom evaluation—often reliant on subjective observations—fall short in scalability, objectivity, and timeliness. This paper proposes a real-time automatic scoring and feedback system for intelligent educational platforms, powered by the Vision Mamba architecture. Vision Mamba—a novel and lightweight spatio-temporal modeling backbone—is integrated with an enhanced classification module, the Augmented Prototypical Network (APN) to form the VMAPN framework, which enables accurate behavior recognition under small-sample conditions. The system analyzes classroom video streams to identify student behaviors such as attentiveness, participation, and posture, and generates adaptive feedback for both students and teachers. Real-time feedback is delivered through an interactive interface, while backend analytics leverage machine learning techniques to monitor learning engagement and evaluate teaching effectiveness. Furthermore, the proposed Teaching Evaluation Model (TEM) classifies student behaviors into four categories—positive, relatively positive, neutral, and negative—to derive objective teaching effect scores. Experimental results on THUMOS 2014 and ActivityNet v1.3 datasets validate the model's predictive performance, achieving mAP scores of 64.7% at IoU 0.5 on THUMOS 2014 and 72.3% at IoU 0.5 on ActivityNet v1.3, representing improvements of 6.2% and 4.8%, respectively, over baseline methods.