Student-YOLOv8: A Privacy-Preserving Deep Learning Framework for Real-Time Student Detection and Automated Attendance Monitoring in Bangladeshi Classroom Environments

Abstract

Automatically tracking attendance in crowded university classrooms remains a major challenge. Currently, most schools rely on manual roll calls, which waste valuable class time and make it incredibly easy for students to skip class and have buddy fake their attendance. This paper presents Student-YOLOv8, a privacy-preserving deep learning framework built on the YOLOv8s architecture for real-time student detection and automatic headcount-based attendance estimate. Unlike biometric or facial-recognition systems, Student-YOLOv8 treats each student as a generic object, eliminating biometric data collection. The DIU-Student dataset-the first annotated classroom image benchmark from a Bangladeshi university-comprises 788 original images augmented to 3,320 samples using Mosaic, MixUp and geometric transformations. Trained for 300 epochs on an NVIDIA T4 GPU, the model achieves precision of 0.768, recall of 0.739, F1-score of 0.753, and mAP@0.5 of 0.737. The system gives per-frame attendance counts without any identity resolution, offering a scalable, ethically compliant and institutionally deployable solution for educational institutions across South Asia.