Examinations are typically used by educational institutions to assess students’ strengths and weaknesses. Unfortunately, exam malpractices like cheating and other forms of academic integrity violations continue to present a serious challenge to the evaluation framework because it seeks to provide a trustworthy assessment. Existing methods involving human invigilators have limitations, as they must be physically present in examination settings and cannot monitor all students who take an exam while successfully ensuring integrity. With the developments in artificial intelligence (AI) and computer vision, we now have novel possibilities to develop methods for detecting students who engage in cheating. This paper presents a practical, real-time detection system based on computer vision techniques for detecting cheating in examination halls. The system utilizes two primary methods: The first method is YOLOv8, a top-of-the-line object detection model, where the model is used to detect students in video footage in real time. After detecting the students, the second aspect of the detection process is to apply pose estimation to extract key points of the detected students. For the first time, this paper proposes to measure angles from the geometry of the key points of detected students by constructing two triangles using the distance from the tip of the nose to both eyes, and the distance from the tip of the nose to both ears; one triangle is sized from the distance to the eyes, and the other triangle contains the measurements to their ears. By continually calculating these angles, it is possible to derive each student’s facial pose. A dynamic threshold is calculated and updated for each frame to better represent the body position in real time. When the left or right angle pass that threshold, it is flagged as suspicious behavior indicating cheating. All detected cheating instances, including duration, timestamps, and captured images, are logged automatically in an Excel file stored on Google Drive. The proposed study presents a computationally cheap approach that does not utilize a GPU or additional computational aspects in any capacity. This implementation is affordable and has higher accuracy than all of those mentioned in prior studies. Analyzing data from exam halls indicated that the proposed system reached 96.18% accuracy and 96.2% precision.
Waleed et al. (Tue,) studied this question.