Adaptive selection of model histograms in block-based background subtraction

Abstract

An adaptive block-based background modelling technique is proposed whereby the optimal number of model histograms is selected. The dynamic nature of a background tends to vary the pool of model histograms when capturing all possible scenes. Proposed is a novel method that recursively estimates the model weights, thereby continuously adjusting the number of histograms to robustly capture only the essence of intended objects. The proposed algorithm shows improved and reliable segmentation performance in various environments, including dynamic backgrounds with moving objects and repetitive variation of the pixel value.