Semi-Supervised Anomaly Detection in Electronic-Exam Proctoring Based on Skeleton Similarity

Anomaly detection in surveillance videos, particularly related to human behavior, is crucial for various applications. In electronic exams (e-exams), cheating can be detected using surveillance videos, but previous research mainly focused on predefined patterns, with less attention given to unsupervised methods. This study proposes a semi-supervised skeleton-based approach for abnormal behavior detection in e-exam proctoring videos. The proposed method segments the skeleton-based feature vectors of consecutive frames based on their similarity. The similarity is calculated based on the Euclidean distance between the mean feature vector and the standard deviation of segments. Similar segments of the training set, which is anomaly-free, are then combined to form distinct frame segments used to recognize the normal samples. Then, in the testing phase, a frame is recognized anomaly if it is not similar to any training segment, otherwise, it is considered normal. In addition, to get a better ranking of anomalous frames, we examine assigning a soft anomaly score considering the segment size and the maximum distance between the comparative features. In label assignment, the algorithm correctly categorizes the testing frame based on their similarity to the training segments. In comparison to the state-of-the-art reconstruction-based anomaly detection algorithms, the proposed method outperforms using the area under the ROC curve (AUC) metric. In addition to correctly detecting similar frames and segmenting videos, the advantage of the presented method is to determine the data labels without needing to process all inputs, making it possible for use in online applications.