A Supervised Approach in Background Modelling for Visual Surveillance

In this paper we address the context of visual surveillance in outdoor environments involving the detection of moving objects in the observed scene. In particular, a reliable foreground segmentation, based on a background subtraction approach, is explored. We firstly address the problem arising when small movements of background objects, as trees blowing in the wind, generate false alarms. We propose a background model that uses a supervised training for coping with these situations. In addition, in real outdoor scenes the continuous variations of lighting conditions determine unexpected intensity variations in the background model parameters. We propose a background updating algorithm that work on all the pixels in the background image, even if covered by a foreground object. The experiments have been performed on real image sequences acquired in a real archaeological site.