Background Subtraction in Varying Illuminations Using an Ensemble Based on an Enlarged Feature Set

Abstract

Image sequences with dynamic backgrounds often cause false classification of pixels. In particular, varying illuminations cause significant changes in the representation of a scene in different color spaces, which in turn results in the high levels of failure in such conditions. Because mapping to alternate color spaces has largely failed to solve this problem, a solution of using alternate image features is proposed in this paper. In particular, the use of gradient and texture features along with the original color intensities are used in an ensemble of mixture of Gaussians background classifiers. A clear improvement is shown when using this method compared to the Mixture of Gaussians algorithm using only color intensities. In addition, this work shows that performing background subtraction using only gradient magnitude as an image feature performs at a much higher rate in varying illuminations then using color intensities. Results are generated on three separate datasets, each with unique, dynamic, illumination conditions.