Unsupervised Video Anomaly Detection Based on Spatiotemporal Generative Adversarial Network

Video anomaly detection has gained significant attention in the current intelligent surveillance systems. Many existing works have difficulties in dealing with the anomaly localization in the crowded scenes due to the lack of sufficient prior information of the objects of interest during training. To cope with these issues, we propose two novel unsupervised anomaly detection and localization methods, Deep Spatiotemporal Translation Network (DSTN) and Deep Residual Spatiotemporal Translation Network (DR-STN). The DSTN is first designed to deal with pixel-level object localization problems based on Generative Adversarial Network (GAN) and Edge Wrapping (EW). However, DSTN faces a false-positive detection rate which affects the overall performance of the system. Thus, the DR-STN is proposed based on the conditional Generative Adversarial Network (DR-cGAN) model with an Online Hard Negative Mining (OHNM) approach to specifically reduce the false-positive detection results. The model architecture is enhanced from DSTN by fashioning a wider network to effectively learn a mapping from spatial to temporal representations for improving the perceptual quality of synthesized images. Our proposed methods have been extensively evaluated on publicly available benchmarks, including UCSD, UMN, and CUHK Avenue. The proposed DR-STN shows superior results over other state-of-the-art methods both in frame-level and pixel-level evaluations. The average Area Under the Curve (AUC) value of the frame-level evaluation of DR-STN for the three benchmarks is 96.73%. The improvement ratio of AUC in the frame level between DR-STN and state-of-the-art methods is 7.6%.