Abstract

Digitalization has become essential for most educational institutions to measure students' knowledge and abilities through examinations. However, cheating remains a persistent issue, with students using various ways such as exchanging answer sheets, using hidden notes, or giving codes. Human monitoring is often inconsistent and limited by focus. This research proposes a solution for solving the cheating detection problem using computer vision, specifically by monitoring the suspicious behavior of students (physical) during the exam through CCTV. The method used to solve the problem is the YOLO (You Only Look Once) algorithm, comparing three versions—YOLOv5, YOLOv6, and YOLOv7. In this study, the accuracy results for each algorithm variation are 43%, 37%, and 51%, respectively. The existence of imbalanced classes in the dataset is the main factor that affects the model performance. Consequently, an extended experiment is undertaken by adding 10% data to the imbalanced classes. The highest accuracy recorded is 60%, given by YOLOv7, reflecting a noteworthy 9% increase in accuracy.

Concepts :

Access to Document

Citations by Year