Video Anomaly Detection using Generative Adversarial Network

Abstract

Abstract—Automatic abnormal event detection and recognition is the need of today’s surveillance scene because public safety is a growing concern these days. The subject is still open for study in the present due to its utility and complexity. Because every individual has a different definition for abnormality, automatically identifying aberrant events is a task not so easy to accomplish. What is normal in one set of circumstances may be viewed as abnormal in another. In a surveillance film of large crowds, automatic anomaly identification is difficult because of traffic and excessive occlusion. This thesis study attempts to provide the answer for this use case by using machine learning techniques, making it possible to eliminate human resources for monitoring the records of the surveillance system for any type of activity that is unexpected. We are going to develop a new anomaly detection model based on GAN, which can train models to learn how to create a high-dimensional picture space and also learn how to extract the latent space of the video from its context. The residual Autoencoder architecture of the generator is coupled with a two-stream deep convolutional encoder that can realize both temporal and spatial data, and a decoder with multi-stage channel attention. Furthermore, we propose an approach to further enhance the performance of the GAN model by knowledge transfer between different datasets in order to reduce training time and improve the generalization capability of the model. We benchmark our model against the state-of-theart methods in current usage on four benchmark datasets using a variety of assessment metrics. The results of empirical study prove that, as compared to methods in use now, our network performs favourably on the dataset. Index Terms—Anomaly, Generator, Discriminator, Surveillance, Traffic, Detection, GAN (Generative Adversarial Network), spatio-temporal, Convolution.