Unsupervised Video Anomaly Detection Using Memory-Augmented Deep Neural Networks

In this study, an unsupervised video anomaly detection approach using memory-augmented deep neural networks is proposed. The proposed approach contains five main stages: (1) optical flow computing of input video frames; (2) the features of video frames (called query) are extracted by the proposed encoder; (3) in the memory module, the similarities between all memory items and the query are computed by the proposed attention-based structure to generate the new query; (4) video frames and corresponding optical flows are predicted by the decoder; and (5) in terms of PSNRs, the anomaly scores of video frames are computed to determine whether video frames are abnormal. Based on the experimental results obtained in this study, the performance of the proposed approach is better than those of seventeen comparison approaches.