Implementation of YOLOv5 Algorithm for Exam Cheating Movement detection

Abstract

The decline in academic integrity due to cheating during exams has become increasingly relevant, particularly following the shift to online learning systems. The absence of direct supervision in online exams creates opportunities for cheating practices that evade detection by the naked eye. This study addresses this challenge by developing an object detection model for cheating behavior using a deep learning approach based on the YOLOv5 algorithm. The dataset comprised 60 ten-second videos, extracted into 1,200 images representing four suspicious head movement patterns. Each image was manually annotated before training five YOLOv5 variants. Models were evaluated using object detection metrics (precision, recall, and mAP at IoU thresholds 0.5–0.95) and analyzed via confusion matrices. Results indicate that the YOLOv5x variant achieved peak performance, with mAP@0.5:0.95 of 83.06% and perfect classification accuracy across all classes. This demonstrates that an object detection–based approach provides a reliable preliminary solution for monitoring cheating during online exams.