Video Summarization in Detecting Overt Cheating Considering Examinee's Interactions with Scene Objects

With the increasing demand for e-exams and the application of video surveillance, it is a time-consuming and tedious task for proctors to review a large number of exam videos. To address this issue, the suggestion is to automatically summarize the videos by considering suspected cheating instances for monitoring purposes. This study proposes a novel approach that leverages deep neural networks to analyze the examinee''s interactions with objects within the scene, with an emphasis on video summarization and efficient processing to identify overt instances of cheating. Overt cheating in this context involves physical interactions with objects in the examinee''s environment, such as referring to notes or using mobile devices, to gain an unfair advantage over fellow examinees. The proposed method comprises three main stages: preprocessing, extracting the examinee''s activities, and detecting unauthorized objects in the scene and their potential use for cheating. Initially, the proposed method filters out frames without the examinee''s movements by comparing the similarity of consecutive frames. Subsequently, the examinee''s movements are identified by analyzing changes in the body skeleton. Finally, suspected cheating scenes are detected by evaluating the interactions between the examinee''s body part movements and unauthorized objects present in the scene. The proposed method offers several advantages, including high processing speed, robustness against background changes, and its ability to operate effectively without prior training. Experimental results, obtained from a dataset of 60 hours of exam videos collected from 120 undergraduate students participating in various assessments, demonstrate that the proposed method successfully detects all instances of overt cheating and suspected cheating while reducing the overall video length by 70%, significantly reducing the time required for manual review.