Real-Time Student Classroom Action Recognition: Based on Improved YOLOv8n
Weiwei Xiang, Haoyu Lei, Liming Ding, Lingfang Zhou · 2023
Real-time student classroom action recognition allows the creation of a closed-loop teaching system and efficient teaching methods. To address the issue of a lack of open dataset in the SCAR re- search field, we proposed a challenging dataset SCAR-dataset with a total of 7194 samples, which was collected in college classrooms and after exploratory experiments. The dataset were defined into six categories: focused, head_drop, bored_distraction, looking_around, sleeping, and stand_up. We benchmarked the SCAR dataset for real-time deployment using YOLOv5, YOLOv7, and YOLOv8 algorithms, which served as the foundation for subsequent classroom concentration analysis research. By combining depth-separable convolution, a lightweight asymmetric detector head LADH was proposed. Experiments show that improving the Yolov8n model increases the map value by 2.5 points while don't in- creasing the number of parameters and Gflops.