Real-time classroom student behavior detection based on improved YOLOv8s

The learning capacity of students is significantly influenced by the quality of instruction they receive in the classroom. With the rapid advancement of behavior detection technology, identifying classroom behaviors of students is becoming increasingly common in educational settings. However, the field still faces specific challenges, primarily concerning the accuracy of identifying student behaviors within complex and variable classroom environments, as well as the real-time capabilities of detection algorithms. To address these challenges, we propose an efficient and straightforward algorithm based on the YOLO architecture. A Multi-scale Large Kernel Convolution Module (MLKCM) has been designed to capture feature information across various dimensions through multi-axis pooling, achieving adaptive receptive fields and effectively capturing multi-scale features. This design enhances the network's sensitivity to feature information by incorporating convolution kernels of varying sizes. Subsequently, we introduce a Progressive Feature Optimization Module (PFOM) to segment the channel dimension of the input feature map. This module integrates feature refinement blocks progressively, which not only preserve the refined features but also efficiently aggregate both local and global information. Finally, we conducted comprehensive experiments using the SCB-Dataset3-S and SCB-Dataset3-U datasets. The results demonstrated mean Average Precision (mAP) values of 76.5% and 95.0%, respectively, surpassing other commonly used detection techniques. Additionally, the effectiveness of our approach was validated through ablation studies and visualization of the detection outcomes.