Dynamic background estimation and complementary learning for pixel-wise foreground/background segmentation

Change and motion detection plays a basic and guiding role in surveillance video analysis. Since most outdoor surveillance videos are taken in native and complex environments, these “static” backgrounds change in some unknown patterns, which make perfect foreground extraction very difficult. This paper presents two universal modifications for pixel-wise foreground/background segmentation: dynamic background estimation and complementary learning. These two modifications are embedded in three classic background subtraction algorithms: probability based background subtraction (Gaussian mixture model, GMM), sample based background subtraction (visual background extractor, ViBe) and code words based background subtraction (code book, CB). Experiments on several popular public datasets prove the effectiveness and real-time performance of the proposed method. Both GMM and CB with the proposed modifications have better performance than the original versions. Especially, ViBe with the modifications outperforms some state-of-art algorithms presented on the CHANGEDETECTION website.