A Deep Learning-Based Method for Classroom Crowd Counting and Localization

In order to count the students’ seating distribution and attendance in offline classroom, which is a better response to the students’ learning and teaching situation. Based on the deep learning method, we propose a crowd localization and counting model for students’ seating area in offline classroom. Firstly, we choose YOLOv8 to improve it, adding the SENet attention module after the backbone network reinforce the role of important channels and speed up model learning, designing a simple and efficient feature fusion method, using the anchor size and the number of detected heads which are more suitable for classroom scenarios and compute the overall loss of the model by using the loss of confidence and the loss of regression of prediction frames. Enhancement methods with Mosaic and cutout data to increase the generalization ability of the model. The improved network achieved 95.405% precision, 92.808% recall and 96.159% mAP on SCUT-HEAD Dataset and C University Dataset.