ABSTRACT

       This paper presents a robust foreground object detection algorithm which can counter the effects of   illumination changes and noise , and thus providing an optimal  choice for intelligent video surveillance systems using static cameras. An Online Expectation Maximization (E-M) algorithm is used in combination with a spherical K-means clustering method for accurate updation of gaussian mixture models when there are changes due to illuminations. The results of the former step is enhanced by the linear RGB color feature of reflection radiance from object surfaces under different environmental illuminations. A statistical framework is used for the foreground object detection. Noise at this stage is  reduced further using  a Bayesian iterative decision-making technique. Various comparative  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.

Keywords: Bayesian, ExpectationMaximization, Guassian mixture models, reflection radiance, robustness