Title:Abnormal Event Detection in Urban Surveillance Videos Using GAN and Transfer Learning

Abstract:Abnormal event detection (AED) in urban surveillance videos has multiple challenges. Unlike other computer vision problems, the AED is not solely dependent on the content of frames. It also depends on the appearance of the objects and their movements in the scene. Various methods have been proposed to address the AED problem. Among those, deep learning based methods show the best results. This paper is based on deep learning methods and provides an effective way to detect and locate abnormal events in videos by handling spatio temporal data. This paper uses generative adversarial networks (GANs) and performs transfer learning algorithms on pre trained convolutional neural network (CNN) which result in an accurate and efficient model. The efficiency of the model is further improved by processing the optical flow information of the video. This paper runs experiments on two benchmark datasets for AED problem (UCSD Peds1 and UCSD Peds2) and compares the results with other previous methods. The comparisons are based on various criteria such as area under curve (AUC) and true positive rate (TPR). Experimental results show that the proposed method can effectively detect and locate abnormal events in crowd scenes.