A vision-based behavioural monitoring framework towards trustworthy AI proctoring for online assessments

This research paper introduces a comprehensive approach to accurate head pose estimation, combining face detection, facial landmark detection, and 3D pose estimation techniques. The methodology holds promise for applications in human-computer interaction, driver monitoring, and security surveillance. The first segment focuses on creating a robust face detection model, achieving an accuracy of 94% with high precision, recall, and F1-score values. The second segment employs a pre-trained Tensorflow CNN model for facial landmark detection, achieving an F1-score of 0.89, with a mean localization error of 1.5 pixels. The final segment utilizes mathematical techniques for precise 3D pose estimation, demonstrating low mean absolute errors for pitch and yaw (2.1 degrees). The integration of these techniques enables accurate real-time head pose estimations with a mean absolute error of 3.2 degrees. Author clarify mentioned in this paper that experiments validate only behavioural monitoring modules. While challenges exist in extreme conditions, future work could explore depth sensor integration for enhanced accuracy and application-specific optimizations. This approach presents a valuable tool with wide-ranging potential across various domains.