基于CA-YOLOv9网络的实时全景多尺度课堂行为识别

Real-time Panoramic Multi-scale Classroom Behaviors Recognition Based on CA-YOLOv9 Network

随着人工智能的发展和“智慧课堂”概念的兴起,课堂行为智能化识别成为研究的热点。目前,国内外研究多采用数个学生或教室的局部影像,而对于学生人数密集、尺度变化范围大且存在大量物体遮挡的全景教室图像实时检测鲜有涉及。为此,文章基于YOLOv9网络,加入CA模块,构建了CA-YOLOv9网络;之后,通过结构分析实验、消融实验和对比实验,得到了CA-YOLOv9网络的最佳结构,并验证了其识别性能;最后,将训练好的CA-YOLOv9网络应用于全景多尺度课堂行为识别,证明了该网络能在不降低推理速度的同时提升检测精度,初步验证了该网络在智慧课堂中实时应用的可行性。文章的研究可为及时了解学生的学习状态和教师教学方法的有效性提供依据,有助于推动人工智能与教育教学的深度融合。

With the development of artificial intelligence and the rise of“smart classroom”concept,the intelligent recognition of classroom behavior has become a research focus.At present,local images of several students or classrooms are mostly used in domestic and foreign studies,but the real-time detection of panoramic classroom images with densely populated students,a wide range of scale changes and a large number of object occlusions is rarely involved.Therefore,based on CA-YOLOv9 network,this paper added coordinate attention(CA)module,and constructed CA-YOLOv9 network.Then,the optimal structure of CA-YOLOv9 network was obtained through structural analysis experiment,ablation experimentand comparison experiment,and further verified its recognition property.Finally,the trained CA-YOLOv9 network was applied to panoramic multi-scale classroom behaviors recognition,proving that the network can improve the detection accuracy without decreasing the inference speed,and preliminarily verified the feasibility of the network’s real-time application in a smart classroom.The research of this paper could provide the basis for timely understanding students’learning status and the effectiveness of teachers’teaching methods,and help to promote the deep integration of artificial intelligence and education and teaching.