Lighter Student Engagement Recognition in a Classroom Environment Using Skeletal Keypoints
Student retention is crucial for educational institutions, influencing reputation, finances, and ranking metrics. Engagement, reflecting a student’s connection, interest, and effort, plays a vital role in learning, fostering critical thinking, and supporting retention. Recent advancements use students’ poses to predict engagement, providing valuable insights without disrupting the teaching-learning dynamic. The prevailing research trend leans toward employing multi-modal approaches, such as a combination of pose detection with object detection. However, current methods use out-of-date object detection methods and manual dataset creation, which is cumbersome, requiring thousands of manually annotated data. This problem is then addressed by proposing a novel method of student engagement detection, using a combination of You Only Look Once Version 8 Mini (YOLOv8m) and MediaPipe, as state-of-the-art alternatives to improve both object detection and human pose estimation. The (YOLOv8+ MediaPipe) method surpasses the baseline (YOLOv4+ OpenPose) with higher accuracy (0.70 vs. 0.41) and lower cross-entropy loss (0.40 vs. 0.60), confirmed by a statistically significant paired t-test. It also exhibits a remarkable speed advantage, being around 18.7 times faster in pose detection data collection rates than the baseline. Despite not being designed for it, the proposed method achieves multiple keypoint detection, matching the baseline’s amount.
Authors:
Gabriel Asael Tarigan, Gregorius Natanael Elwirehardja, Kuncahyo Setyo Nugroho, Bens Pardamean
IAENG International Journal of Computer Science