Video anomaly detection using diverse motion-conditioned adversarial predictive network.

Video anomaly detection is always formulated as frame prediction task which only learned on normal data and detects deviations as anomalies. However, previous methods lack sufficient spatiotemporal constraints on moving objects, making it difficult to learn compact normal distributions and anomalies near the boundary will be misclassified as normal. Besides, the inadequate exploration of diverse normal patterns results in mode missing and unlearned normal patterns will be misclassified as anomalies. To address these problems, we propose an object-level Diverse Motion-conditioned Adversarial Predictive Network for video anomaly detection which combines conditional variational generation with adversarial learning to mitigate false detection. We design a motion-guided generator that controls the generation process conditioned on optical flows to accurately memorize spatiotemporal correlations of normal data. We employ the diversity regularization strategy which explicitly preserves the recurrent structure of normal data in continuous latent space to ensure full utilization of diverse patterns. Additionally, we combine an input clip with the object it generates to synthesize an anomaly near the boundary, then employ a video discriminator to perceive subtle differences between normal and abnormal data, making them more distinguishable. Extensive experiments conducted on public datasets illustrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]