Robust Cognitive Engagement Detection with High Density and Low-Quality Data in Real-World Classrooms

Student cognitive engagement is defined as the extent to which students are willing and able to take on the learning task at hand. It can be inferred by analyzing students’ behavioral data during the learning process. However, existing analysis methods often suffer from low accuracy, poor real-time performance, and inadequate consideration of sample quality differences. In high-density classroom settings, especially in back rows, current models face challenges in accurately detecting students’ cognitive engagement due to multiscale detection limitations. Therefore, developing an efficient and accurate model for detecting student cognitive engagement is of great significance. This study introduces a new WEVC-YOLOv10s cognitive engagement detection model, proposing an efficient EVC attention mechanism and a W-IoU loss function to improve accuracy. Experimental results show that the improved model achieved 84.7% detection accuracy, 82.5% recall, and 86.6% mAP50, surpassing the traditional YOLO model in multiple evaluation metrics, demonstrating its application potential in intelligent education systems.