An Advanced Codebook Background Model Using Confidence and Membership Function

This paper introduces an advanced codebook model for foreground-background segmentation. The improvement is two-fold. First, to cope with changing global lumination, we develop the conventional codebook with confidence functions. Similarities of both brightness and normalized color-vector are integrated confidence-weightedly to form overall similarity. A designed codeword-update progress also contributes to the stable performance. Besides, by introducing a membership function based measurement, the threshold of overall similarity is able to adjust itself with statistical properties of the given video. It makes the model rather robust. A thorough evaluation is performed on the Wallflower dataset. Qualitative and quantitative results and comparisons with other approaches justify the model.