Visual surveillance using background model image generated by GAN

Visual surveillance requires robust foreground and background separation capabilities in various environments. Although various traditional algorithms based on background subtraction methods have been proposed, problems such as hard shadows, camouflage, and ghost effects remain. Recently, deep learning-based foreground detection methods have been proposed. Deep learning-based methods outperform traditional algorithms in various unmanned surveillance datasets. However, even deep learning-based methods show insufficient generalization ability in certain datasets. For data that have not been trained, a number of errors are detected. Even among deep learning-based methods, there are methods that show higher generalization ability by using a background image. In this paper, we propose a method of using GAN to generate background images.