Modelling of the background (“uninteresting parts of the scene”), and of the foreground, play important roles in the tasks of visual detection and tracking of objects. This paper presents an effective and adaptive background modelling method for detectin

2.3): to cope with shadows, normalized colour is used (sections 2.3.1and 2.3.2); treatment of isolated holes in foreground objects (section 2.3.3); the setting of the essential threshold controlling the notion of consensus (section 2.3.4) and the use of a Time Out Map to control which pixels are added to update the background samples (section 2.3.5). These components are put together into an overall background modeling framework (section 2.3.6).

2.1 Related Work of Background Modelling

Numerous background modelling studies have appeared in the literature in recent years, [3, 10, 11, 17, 20, 22, 23, 26, 27, 29]. A simple background model usually assumes that the background pixels are static over time. The foreground objects can then be obtained by subtracting the current frame from the background image. Realistically, background models have to allow for a distribution of background pixel values as lighting changes etc. For example, to capture the allowed variation in background values, W4 [10] models the background by maximum and minimum intensity values, and the maximum intensity difference between consecutive frames in the training stage. Other techniques assume a statistical model: Pfinder [29] assumes that the pixels over a time window at a particular image location are single Gaussian distributed. Although these methods can deal with small or gradual changes in the background and they work well if the background includes only a static scene, they may fail when background pixels are multi-modal distributed (e.g., waving trees) or widely dispersed in intensity.

Several methods have been proposed to deal with multi-modal distributed background pixels. Wallflower [27] employs a linear Wiener filter to learn and predict background changes. Wallflower works well for periodically changing pixels. However, when the background pixels change dramatically or the movement of those background pixels are less periodical, Wallflower is less effective in learning and predicting background changes.

5