Research on Multi-modal Dense Scene Individual Behavior Detection Based on Improved YOLOv13

With the rapid development of smart education, traditional classroom teaching evaluation methods face severe challenges in real-time performance, objectivity, and efficiency. Based on the YOLOv13 algorithm, this paper proposes an improved model integrating collaborative dual attention mechanisms. By designing a parallel dual-path architecture, the model achieves synergistic effects between channel attention and spatial attention, while innovatively adopting a shared Query-Key mechanism that significantly enhances feature discriminability while maintaining high computational efficiency. The study constructs a dataset containing six typical classroom behaviors (raising hands, reading, writing, using phones, looking down, leaning on desks). Experimental results demonstrate that the improved model achieves significant performance improvement on the self-built dataset, reaching an mAP@0.5 of 82.7%, a 6.4% improvement over the baseline model. This research not only provides an effective automated solution for classroom behavior analysis but also offers new insights into applying attention mechanisms in real-time visual tasks.