On-Line Learning of Background Appearance Changes for Robust Background Subtraction

We present a background subtraction approach aimed at efficiency and robustness to common sources of disturbance such as illumination changes, camera gain and exposure variations, noise. The novelty relies in trying to learn, at each new frame, a model of the background intensity changes currently yielded by the sources of disturbance. Based on the observation that such changes are highly correlated across large portions of the image, a unique frame-wise model is used, which consists in the bivariate probability density function of background and frame intensity at a pixel. After the non-parametric estimation of the model through the bivariate histogram, changes are detected by thresholding the histogram entries. Experimental results prove that the approach is state-of-the-art in challenging sequences characterized by sources of disturbance yielding sudden and strong background appearance changes.