Abstract: In unsupervised video anomaly detection tasks,deep autoencoders are typically trained on datasets containing only normal events and use reconstruction(prediction) error to identify anomalous frames.However,this assumption does not always true in practice because sometimes autoencoders can reconstruct(predict) anomalous events well,leading to false alarms.To address this issue,this paper proposes a video anomaly detection method based on dual discriminators and pseudo video generation,which enhances the generation model’s prediction capability of normal frames and suppresses its prediction capability of pseudo video frames through adversarial training between the discriminator and the generator.Moreover,the introduction of coordinated attention in the generation model further improves its detection performance.Additionally,by predicting intermediate frames instead of future frames in previous methods,the model can learn forward and backward motion information,which further enhances its detection performance.Experimental results on the publicly available datasets UCSD Ped2 and CUHK Avenue demonstrate that the proposed method achieves AUC values of 98.6% and 85.9%,respectively,outperforming other video anomaly detection methods significantly.

Key words: Video anomaly detection, Deep learning, Generative adversarial network, Pseudo-video, Prediction