Background estimation for dynamic video scenes

Background estimation in video is an important process in many surveillance applications. Usual methods model the temporal variation of background pixels, assumed small and repetitive. This assumption is not valid with rapid, strong luminance and color changes as well as small motion. Moreover, initialization and update of such models require the observation of many frames. This paper presents a novel approach where variations are handled spatially in a global and unified manner, allowing robustness to drastic changes, and a fast update of the estimation. The main aspects of the proposed method are the optimal estimation of observed background variations, an image representation as a field of color distributions, the handling of multiple hypotheses for robustness to foreground objects and the fast computation of local statistics. Results for public transportation vehicles at a bus stop, as well as for synthetic data are presented