Traditional examination room monitoring systems face challenges such as low recognition rates and insufficient real-time performance. Existing behavior recognition methods are further limited by environmental factors, including occlusion, small targets, and subtle feature differences. To address these issues, an improved YOLO11n-based model (YOLO11n_BSS) is proposed. This model introduces a Bi-Level Routing Attention (BRA) to address inherent occlusion problems in examination rooms, focuses on enhancing small target detection capabilities by integrating StripPooling with the C3k2 module, and incorporates Single-Head Self-Attention (SHSA) into the C2PSA module for optimizing fine-grained feature extraction. Experimental results demonstrate that compared to the baseline model, the proposed model achieves a 7.3% increase in precision (P), a 4.7% increase in recall (R), a 2.2% increase in mean Average Precision (mAP), and a 3.7% increase in mAP50-95, thereby significantly improving performance.