Abnormal Event Detection and Localization in Visual Surveillance

In this paper, we propose a framework for abnormal event detection and analysis in the field of visual surveillance based on the state-of-the-art deep learning techniques. We train a pair of conditional generative adversarial networks (cGANs) using the normal behavior samples, where one cGAN takes video frames as inputs and generates the corresponding optical flow features. While on the other hand, the other cGANs take optical flow features as inputs and generate the corresponding video frames. By analyzing the differences between the generated frames/optical flow features and the realistic samples, abnormal events can be detected and localized effectively. Moreover, for suspected regions, we adopt the faster RCNN to analyze the abnormal events. Experimental results demonstrate that the proposed framework can detect the abnormal events accurately and efficiently.