Illumination-Robust Foreground Detection in a Video Surveillance System

This paper presents a foreground detection algorithm that is robust against illumination changes and noise, and provides a novel and practical choice for intelligent video surveillance systems using static cameras. This paper first introduces an online expectation-maximization algorithm that is developed from a basic batch version to update Gaussian mixture models in real time. Then, a spherical K-means clustering method is combined to provide a more accurate direction for the update when illumination is unstable. The combination is supported by the linearity of RGB color reflected from object surfaces, which is both theoretically proved by spectral reflection theory and experimentally validated in several observations. Foreground detection is carried out using a statistical framework with regional judgment. Noise in the detection stage is further reduced by a Bayesian iterative decision-making step. The experiments show that the proposed algorithm outcompetes several classical methods on several datasets, both in detection performance and in robustness to perturbations from illumination changes.