Video Anomaly Detection Based on Deep Generative Network

In this paper, we present a framework for the detection of anomalies in video scenes. Both spatial and temporal features extract and learn through the framework. We employ inception modules and residual skip connections inside the framework to make the network learning higher-level features, which we call "multi-scale U-Net". A multi-scale U-Net kept useful features of the image that lost during training caused by the convolution operator. The numbers of training and testing parameters in our framework are reduced while the detection accuracy is still improved. We evaluated the proposed framework on three benchmark datasets: UCSD, CHUK Avenue and ShanghaiTech dataset. Our proposed framework achieved 95.7%, 86.8% and 73.0% in terms of AUC, which surpasses the state-of-art learning-based methods.