Semi-supervised Anomaly Detection and Location Based on Generative Adversarial Network

Abstract

With the development of artificial intelligence, the anomaly detection plays more and more important role in security monitoring field. Because it is difficult to label abnormal data, most of the supervised methods consumed a lot of manpower and obtained low performance and generality. Inspired by this motivation, this paper proposes a semi-supervised method for anomaly detection in video frames based on GAN (Generative Adversarial Network), in which only normal data was used as the training sample. The quality gap between the predicted frame and the ground truth is used as the basis to determine whether it is abnormal. Moreover, the mathematical morphology approach was adopted to locate the anomaly area in the frames. Experiments show that our method can successfully detect abnormal frames in video and can also locate the area where abnormal behavior occurs in frames.