Design and analysis of GAN architecture for anomaly detection

Abstract

In order to maintain public safety, surveillance cameras are progressively being utilized in public spaces like roadways, crossroads, banks, etc. Video footage from these cameras helps avoid criminal or unwanted activities. A large number of people have been employed to monitor the surveillance system in which unexpected events occur once in a while. It also needs manual operations to assess the tape for any unlikely event. This thesis work is meant to provide the initial solution for this use case using machine learning techniques to avoid using human resources in monitoring any anomalous activities in surveillance system recordings. We use video frames from the past and present to identify anomalous activity and predict unprecedented future events. We model a reconstruction-based, One-Class Classifier(OCC) using the Generative Adversarial Network(GAN) architecture to classify a given dynamic image frame as anomalous or normal. The prediction model, with one generator and two discriminators, is trained with only normal samples and the abnormal or anomalous samples are considered outliers. Thus when the model receives an abnormal sample(outlier), it is expected that the reconstructed result would be poor, helping us in detecting the abnormal frame.