A WAD-YOLOv8-based method for classroom student behavior detection

This paper proposes an enhanced YOLOv8 model to address the challenges of complex classroom behavior detection. The model effectively overcomes the limitations of the original YOLO backbone, including the restricted receptive field and insufficient multi-scale feature learning due to fixed convolutional kernels. We introduce a novel CA-C2f module, enabling more comprehensive fusion and adjustment of the receptive field. Additionally, we propose the attention-based 2DPE-MHA module, which enhances the model's ability to capture long-range dependencies, thereby improving detection performance for multi-scale, occluded, and small targets. The model also incorporates a dynamic sampling factor, Dysample, which selectively focuses on regions with rich details, alleviating the potential loss of detail associated with traditional fixed sampling strategies and further boosting model performance. Experimental results demonstrate that the proposed model outperforms existing methods on public datasets such as SCB, SCB2, SCB-S, and SCB-U, with mAP@0.5 improvements of 2.2%, 3.3%, 5.5%, 18.7%, respectively, and mAP@0.5:0.95 improvements of 3.2%, 2.3%, 3.5%, 14.8%. Moreover, the model maintains real-time inference capabilities that surpass those of other object detection models. The application of this model can assist student work managers in effectively monitoring classroom behavior.