Multi-Scale Classroom Behavior Recognition Study Based on CA-YOLOv7

The identification and study of classroom behaviors is very important for evaluating the quality of teachers' teaching as well as students' engagement in class. Compared to elementary, middle and high school classrooms, in a university lecture setting, teachers have more teaching tools, students carry electronic devices and have more diverse body movements. Therefore, we constructed a dataset of dense crowds from real classroom videos. Meanwhile, for classroom behavior recognition in a dense classroom and multi-scale student group environment, general deep learning target detection models cannot well effectively incorporate global information such as the overall human profile. In this paper, based on YOLOv7 network, the CA (Coordinate Attention) attention mechanism module is added to improve the detection ability in the case of occlusion as well as the absence of the human body picture through the position information and the channel attention mechanism. The experimental results show that the model achieves a mean average precision (mAP) value of 71.5% on the self-built RUSCB-dataset classroom behavior dataset, which is a 1.3% improvement over the baseline model, and the inference time is 56ms, proving that the model enhances detection accuracy without compromising inference speed. This enables effective detection of students' classroom behaviors in real classroom environments.