Abstract

cl We propose a joint foreground-background mixture prmodel (FBM) that simultaneously performs background meestimation and motion segmentation in complex dynamic DTscenes. Our FBM consist of a set of location-specific dypunamic texture (DT) components, for modeling local backthground motion, and set of global DT components, for modtaeling consistent foreground motion. We derive an EM albagorithm for estimating the parameters of the FBM. We also apply spatial constraints to the FBM using an Markov ranmidom field grid, and derive a corresponding variational aptaproximation for inference. Unlike existing approaches to Fibackground subtraction, our FBM does not require a manbaually selected threshold or a separate training video. Unanlike existing motion segmentation techniques, our FBM can cosegment foreground motions over complex background with fimixed motions, and detect stopped objects. Since most dysenamic scene datasets only contain videos with a single forejoground object over a simple background, we develop a new pechallenging dataset with multiple foreground objects over ticomplex dynamic backgrounds. In experiments, we show olthat jointly modeling the background and foreground segwiments with FBM yields significant improvements in accunoracy on both background estimation and motion segmentagrtion, compared to state-of-the-art methods. ob