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

Keywords: Background modelling, background subtraction, sample consensus, visual tracking, segmentation, foreground appearance modelling, occlusion.

1. Introduction

Background modelling is an important and fundamental part for many computer vision applications such as real-time tracking [2, 20, 21, 25, 26], video/traffic surveillance [9, 10] and human-machine interface [23, 29]. After the background is modelled, one commonly performs “background subtraction” to differentiate foreground objects (those parts are of interest to track or recognize) from the background pixels. The result of background modelling significantly affects the final performance of these applications.

Generally speaking, a good background model should be able to achieve the following desirable properties:

accurate in shape detection (i.e., the model should be able to ignore shadow, highlight, etc.); reliable in different light conditions (such as a light switched on/off, gradual illumination changes) and to the movement of background objects (e.g., if a background object is moved, that object should not be labelled as a foreground object);

flexible to different scenarios (including both indoor and outdoor scenes);

robust to different models of the background (i.e., a time series of observation at a background pixel can be either uni-modal or multi-modal distributed) and robust in the training stage even if foreground objects exist in all training examples;

accurate despite camouflage (e.g., if a foreground object has similar color to the background) and foreground aperture (if a homogeneously colored object moves, many of the interior pixels of the object may not be detected as moving);

2