AI-Enabled Learning Behavior Analysis in Smart Classrooms Based on DINOv3

With the development of smart classroom platforms, the acquisition, analysis and feedback of learning data have become key topics in educational technology research. Compared to traditional assessment methods such as post-class tests and questionnaires, learning behavior detection can capture student engagement, providing more direct evidence to support processbased evaluation. However, real-world classroom scenarios commonly present challenges such as small-scale targets, subtle behavioral differences, frequent occlusions, and complex backgrounds. These issues pose significant challenges for automated behavior detection. In this study, we propose a learning behavior analysis framework based on DINOv3, designed for smart classroom scenarios. The framework employs a DINOv3 visual backbone to extract robust semantic features and integrates a transformer-based detection architecture inspired by DEIMv2 to locate and classify student behaviors. This paper analyzes typical classroom behaviors in grades K-12 and universities, and further explores its application value in student engagement analysis and intelligent feedback support. The research demonstrates that incorporating advanced visual representation models into classroom behavior analysis provides a viable approach to AIsupported interpretable classroom teaching evaluation.