VALD-GAN: video anomaly detection using latent discriminator augmented GAN

The most crucial and difficult challenge for intelligent video surveillance is to identify anomalies in a video that comprises anomalous behavior or occurrences. The ambiguous definition of the anomaly makes the detection of it a challenging task. Inspired by the wide adoption of generative adversarial networks (GANs), we proposed video anomaly detection using latent discriminator augmented GAN (VALD-GAN), which combines the representation power of GANs with a novel latent discriminator framework to make the latent space follow a pre-defined distribution. We show through our experimental results that the proposed method significantly increases the anomaly discrimination capability of the model. VALD-GAN achieves an AUC and EER score of 97.98, 6.0% on UCSD Peds1, 97.74, 7.01% on UCSD Peds2, and 91.03, 9.04% on CUHK Avenue dataset, respectively. Also, it is able to detect 62 out of a total of 66 anomalous events with 4 as false alarms and 19 out of a total of 19 with 1 false alarm from Subway Entrance and Exit video datasets, respectively.