Student Pose Detection Algorithm for the Classroom Environment Based on Improved YOLOv7

The field of smart education is experiencing rapid development in artificial intelligence and deep learning technologies. Researchers are increasingly focusing on classroom intelligence and informatization. The use of classroom monitoring has become popular, making it valuable to automatically gather information about students' learning status through cameras. While pose estimation techniques have been extensively studied in recent years, applying these methods in different school and classroom scenarios with diverse camera positions remains challenging. Unlike pose estimation methods, we consider pose detection of students in classrooms as a distinct problem of object detection. However, the monitoring data in the classroom scene presents challenges, such as similar categories, small object sizes, and severe occlusion. Achieving a balance between accuracy and speed in pose detection is often difficult. To address these problems, we propose an attitude detection algorithm that combines feature fusion with a one-stage object detection framework. Firstly, we utilize the YOLOv7 algorithm framework to preprocess the image and improve the data before feeding it into the network. Secondly, we introduce a feature extraction module that combines self-attention and convolution into the network, allowing the algorithm to quickly extract and prioritize features. Finally, we incorporate a feature fusion enhancement module multiple times into the network structure to obtain multi-scale features and enhance the model’s learning ability. Experimental results demonstrate that our proposed method exhibits a slight advantage over other excellent object detection methods when applied to real classroom pose datasets.