STUEFF-YOLOv5s: A Lightweight Method for Deep Identification of Students’ Behavior in the Classroom
Xiaolin Fan, Binghao Fan · 2024
Identifying students' classroom behaviors through classroom activity videos plays an important role in helping teachers improve teaching quality and focus on students' healthy growth and learning habits. Existing student behavior recognition methods mainly focus on the behavior of a single student, with low performance and efficiency. In order to identify the behaviors of multiple students in a classroom simultaneously, we propose a fast and effective solution STUEFF-YOLOv5s, which is an improved YOLOv5s based on the fusion of an improved lightweight network EfficienctCBAM and a lightweight arithmetic feature pyramid. First, the improved lightweight network EfficienctCBAM replaces the original backbone network to achieve better student behavioral feature extraction. Then, CARAFE is fused into the original feature pyramid structure of YOLOv5, which improves the retention of student behavioral details, such as subtle differences between reading students and writing students, while enhancing spatial consistency. Finally, the experimental results show that the proposed solution can detect eight common student behaviors in the classroom, including listening, writing, reading, standing, head turning, hand raising, group discussion, and teacher instruction. It has less number of parameters and better performance compared to tiny target detection algorithms such as YOLOv4.