To address the problem of feature blurring and accuracy degradation caused by occlusion in real classroom environments, an improved YOLOv8 model for student classroom behavior recognition, OA-YOLOv8, was designed. An auxiliary occlusion prediction head (OPH) was introduced in parallel to the YOLOv8s model's head structure. Through an online data augmentation strategy, varying degrees of occlusion were dynamically simulated and proxy occlusion labels were generated to supervise the learning of the OPH module. The model then predicted the occlusion degree as a categorical attribute, alleviating the feature blurring caused by occlusion. Experimental validation was conducted on the SCB classroom behavior dataset. Results show that OA-YOLOv8 achieves 4.6% and 3.2% higher accuracy (mAP@0.5 and mAP@0.5:0.95) than the baseline YOLOv8s. Performance improved by 6.2% and 8.8% in moderate and heavy occlusion scenarios, respectively. The average precision improved by 4.9% when distinguishing fine-grained behaviors with similar visual features. This model effectively improves the accuracy and robustness of student classroom behavior recognition in complex occlusion environments, and provides more reliable technical support for personalized teaching analysis and feedback in the construction of smart campuses.