AI proctoring for offline examinations with 2-Longitudinal-Stream Convolutional Neural Networks

In recent years, Artificial Intelligence (AI)-assisted proctoring techniques have developed rapidly, but most are for online exams. On the contrary, this study aims to explore the AI-based visual proctoring approach through surveillance cameras in real-world offline examination rooms (usually classrooms) to assist proctoring, with contributions from three perspectives: First, to proctor in the offline examination rooms, a visual assessment system is introduced that can be applied in the majority of various classrooms for proctoring after corresponding model training. Second, the new proposed 2-Longitudinal-Stream Convolutional Neural Networks model can adequately adapt to the large-scale examination room environment (i.e., the surveillance cameras are too far away from some examinees) with different lighting conditions. In the third aspect, this study proposes a dimensionality reduction technique called Mini-Batch Cutting for video data capable of converting video information into almost lossless image information with a corresponding data preprocessing method. Compared to existing studies, the proposed method has lower computational requirements and is easier to use in all exam rooms simultaneously. Finally, the study findings demonstrate that the proposed proctoring system has a better detection effect than humans in real-world large-scale offline examination environment tests