Traditional classroom teaching evaluations often rely on subjective feedback and post hoc academic metrics, which limits their objectivity and timeliness. To address these limitations, this study proposes YOLOv11_BWC, a real-time object detection framework for student behavior detection, classroom attention assessment, and teaching quality evaluation in intelligent education. YOLOv11_BWC integrates a BiFormer module to improve occlusion robustness in complex classroom scenes, a Wise-IoU v3 loss to enhance multi-scale localization, and a lightweight Shuffle Attention mechanism to improve computational efficiency. Using real classroom videos from 257 students, three posture-related behaviors, namely head-up, head-down, and sprawled posture, were annotated, augmented from 840 to 1,326 images, and used for model training and evaluation. On the classroom behavior dataset, YOLOv11_BWC achieved 0.9672 mAP@0.5, 0.7594 mAP@0.5:0.95, 0.931 recall, and 136 FPS, outperforming YOLO26, YOLOv8, YOLO11, RT-DETR, and YOLOv12 in the overall balance between detection accuracy and real-time performance. Additional Jetson Orin Nano testing and POCO/SCB-Dataset3 validation further supported its embedded deployment efficiency and cross-domain robustness. Heatmap visualizations verified more focused feature responses, and temporal analysis revealed an exploratory classroom attention pattern of high initial attention, mid-class decline, and final resurgence. The framework provides objective, non-invasive, and fine-grained behavioral evidence for instructional feedback and dynamic classroom quality assessment.