Subtract Background (OpenCV)

How it fits

Typical backends

Caveats

• AccuracyThe background model needs to observe several hundred frames before its output stabilizes; foreground extracted during this warm-up is noisy, and the warm-up length grows with history.
• AccuracyA panning or shaking camera and sudden lighting shifts both flood the model with apparent change, saturating the foreground until it re-adapts; the model assumes a largely static scene.
• State lifetimeThe model carries learned background statistics across every tick, so a restart discards it and the first frames after restart are unreliable again.
• FallbackWhen shadow marking is enabled, shadow pixels are written as mid-gray rather than dropped; treating every non-zero pixel as motion silently folds shadows into the foreground unless that behaviour is turned off.
• Parameter interactionFreezing adaptation by setting learning_rate to zero before the model has seen history frames locks in an under-trained background and produces persistent false foreground.