An Adaptive Learning Rate Method for Improving Adaptability of Background Models

Many popular background modeling (BGM) methods update the background model parameters using an exponentially weighted moving average (EWMA) with fixed learning rates, which cannot adapt to diverse surveillance scenes. In this letter, we propose a statistical method to generate adaptive learning rates for the EWMA-based BGM methods. The method defines a novel way to analyze pixel intensity variations in video sequences and builds an intensity-level migration probability map, which is a recursively updated 2-D lookup table for retrieving adaptive learning rates. Experimental results demonstrate the proposed method can effectively improve the adaptability of the EWMA-based BGM methods across different surveillance scenes.