Context-related video anomaly detection via generative adversarial network

Video anomaly detection is a challenging task because of the scarcity and ambiguity of abnormal event samples in videos. Predicting future content based on continuous video content provides a key clue for anomaly detection. However, most existing methods ignore the video context, which is important for understanding content. To address this issue, we propose a context-related video anomaly detection method combined with a generative adversarial network. Specifically, given a video frame, we first generate its content with some continuous frames before and after this frame by using a two-branch generator network, and then minimize the prediction errors between the generated frames and its ground truth. In addition, to explore the bidirectional relations well, we minimize the content loss between the generated frames and the ground truth in two directions. In the anomaly detection phase, the final anomaly score is calculated by combining the two directional generators obtained in the training phase. Results of experiments on three benchmark datasets demonstrate the effectiveness of the proposed method.