The design and implementation of a student classroom behavior recognition system based on YOLOv8

With the advancement of smart education, traditional classrooms are undergoing a transformation into smart classrooms, and the recognition system for student classroom behavior, which contributes to enhancing teaching efficiency, has emerged as a prominent research focus. Mainstream target detection algorithms encounter several challenges including occlusion, limited availability of open real classroom environments, and scarcity of databases for analyzing student behavior. In this paper, we propose an enhanced model based on YOLOv8 to analyze student behavior. Firstly, we develop an open-source dataset comprising 5000 keyframes sourced from various real classroom videos. Secondly, we incorporate the multi-head self-attention mechanism (MHSA) and Ghost Convolution are into YOLOv8 to improve the recognition precision and increase running speed. Experimental results demonstrate that our proposed method improve the mean Average Precision ([email protected]) value from 81.8% to 84.9%, accuracy rate rises from 77.9% to 81.1%, recall rate improves from 75.4% to 78.2%. Furthermore, the reasoning time is reduced by 16.7%, and the number of parameters is decreased by 23.0%.