Anomaly event detection using generative adversarial network for surveillance videos

Anomalous event detection is advantageous for real-time video surveillance systems in terms of safety and security. Current works mostly run offline and struggle with abnormal event detection in crowded scenes. We propose unsupervised anomaly event detection using Generative Adversarial Network (GAN) with Optical Flow to obtain spatiotemporal features in appearance and motion representations. In training, GAN is used to train only the normal event images to generate their corresponding optical flow image. Hence, in testing, since the model knows only the normal patterns, any unknown events are considered as the anomaly event which can be detected by subtracting the pixels between the generated and the real optical flow images. We implement on the publicly available benchmark datasets and compare with state-of-the-art methods. Experiment results show that our model is effective for anomaly event detection in real-time surveillance videos.