Dominant and rare events detection and localization in video using Generative Adversarial Network

Dominant and rare events detection is one of the most important subjects of image and video analysis field. Due to inaccessibility to all rare events, detecting of them is a challenging task. Today, deep networks are the best tool for video modeling but due to inaccessibility to tagged data of rare data, usual learning of a deep convolutional network is not possible. Due to the success of generative adversarial networks, in this paper an end-to-end deep network based on generative adversarial networks is presented for detecting rare events. This network is competitively trained only by dominant events. To evaluate performance of proposed method, two standard datasets: UCSDped1 and UCSDped2 are utilized. The proposed method can detect rare event with 0.2 and 0.17 equal error rate with the processing speed of 300 frames per second on the mentioned data respectively. In addition to end-to-end structure of the network and its simple train and test phase, this result is comparable to advanced methods results.