Many applications in computer vision use some type of background model as the first stage of video processing in order to separate interesting objects from their surroundings.

For example the surveillance system for a station concourse might typically utilize a background model to locate and highlight people and luggage items whilst ignoring the static structure and fixtures of the building.

The essence of the separation process is firstly to obtain a mathematical model of the true background, and then secondly to find the difference between this model and actual scenes. The difference represents objects which are not part of the background.

A simplistic model might use the pixel-by-pixel intensity average of many images as the background. Such a solution can be satisfactory in scenarios exposed to constant lighting conditions, such as artificially lit indoor or underground environments. However, many systems are required to function in outdoor situations where daylight and the weather are uncontrollable parameters. Rain or snow can obscure the view, wind causes trees to sway, and a scene's appearance may be altered drastically by illumination changes as clouds occlude the sun.
More advanced background models use statistical and adaptive techniques in attempt to deal with such problems. A review of some traditional and contemporary approaches is presented herein.

Return to Conference Programme