Abstract: To address the problems of dense targets, frequent mutual occlusions in student behaviors with classroom scenarios and limited computing power of edge devices, this paper proposes a lightweight YOLOv8-based student classroom behavior recognition method for edge computing, which maintains detection accuracy while significantly reducing model complexity. First, a Neck-DSC lightweight feature fusion structure is proposed. It reconstructs the feature fusion path through depthwise separable convolutions, reducing computational complexity basically without losing representation capability. Secondly, a Dynamic Sparse Attention (DSA) mechanism is designed. With detection confidence as a prior, it adaptively allocates attention computation resources only to potential occluded regions, thereby enhancing the behavioral feature modeling capability in occlusion scenarios. Finally, the INT8 quantization deployment is achieved by integrating TensorRT, further improving the inference efficiency of the model on edge devices. Experiment results on a self-constructed classroom behavior dataset show that the proposed method has a model parameter count of only 2.39×10⁶, presenting a 24.13% reduction compared with the original YOLOv8n, while maintaining mAP at approximately 95%. Ablation studies validate the effectiveness of the Neck-DSC and DSA modules in terms of lightweight capability and occlusion robustness.