Background Subtraction

Background pixels exhibit natural variation: surfaces reflect light at slightly different angles, causing intensity fluctuations; vegetation and water move subtly. A single Gaussian cannot capture this multi-modal behavior. A mixture of Gaussians allows some components to represent stable background (low variance) and others to represent non-foreground variations (low confidence).

The background model is updated continuously: when a pixel matches the background model, that observation is used to adjust the Gaussian components (mean and variance). This incremental learning allows the model to drift gradually toward new illumination conditions. However, sudden changes (like turning on a light) can overwhelm the model until enough frames have been processed.

Shadows are regions where foreground objects block light from the background, darkening that area. Naive background subtraction classifies shadows as foreground because their intensity is different from the background. Shadow detection identifies regions where intensity decreased proportionally (shadow signature) and classifies them as background, improving foreground object boundaries.

Occlusions (a person standing still) gradually become absorbed into the background model because the area matches the model's Gaussian distribution for enough frames. To prevent this, some methods explicitly track persistent foreground regions or use a separate 'shadow' component for long-term occlusion. Re-lighting is handled by model adaptation, though sudden changes still cause false detections until the model adjusts.