Speakers
Description
The rise of academic dishonesty through exam
cheating has created more advanced and complex plagiarism
detection systems. These systems are proprietary; they come
with financial, hardware, privacy, and cloud requirements that
can’t be covered by all educational institutions. An open-source
real-time video plagiarism detection system focused on suspicious
gaze direction and cheating device detection, which can run on
accessible hardware, is proposed. For the gaze direction detection,
the system uses the MediaPipe Face Mesh landmarks, a Kalman
filter, horizontal ratio, vertical ratio, and dynamic thresholds to
obtain an accuracy of 92.4% with a latency of under 100 ms on
our own proposed gaze direction manually annotated dataset. The
object detection part is done by two YOLOv8 models fine-tuned
with two specific datasets to offer an 88.6% smartphone detection
92.5% smartwatch detection with under 200 ms latency. The
system runs only on a CPU with a memory footprint under 820
MB, with over 25 FPS. It only stores the suspicious parts in
the entire video, making it suitable to use on small educational
institution systems and reducing privacy concerns for storage
