Student Behavior Recognition Based on Multi-scale Deformable Graph Convolution via Feature Fusion and Its Application in English Teaching

This paper proposes a multi-scale deformable graph convolutional network (MD-GCN), which integrates spatio-temporal and semantic features to accurately identify students’ behaviors in the classroom, and verifies its immediate feedback value in English teaching. The model takes the classroom monitoring sequence as the input, it first uses lightweight pose estimation to extract joint coordinates and construct a dynamic human body graph. Then it designs multi-scale deformable convolution kernels to adaptively capture the topological changes of fine-grained actions such as raising hands, reading, and whispering, and through the cross-layer feature fusion module, aggregates short-term poses, long-term trajectories, and text semantic information. On the English classroom dataset, the mAP of the proposed method reaches 91.7%, and the inference speed is 38 FPS , which can meet real-time requirements. The teaching experiment shows that after the system pushes the recognition results to the teacher’s end in real time, the response rate of classroom questions increases by 24%, and the duration of students’ concentration increases by 17%. This research provides an expandable artificial intelligence solution for smart intervention in smart classrooms.