Kernel selection for mean shift background tracking in video surveillance

Abstract

Nonparametric kernel density estimation has been successfully used in modeling the background statistics, in video surveillance, due to its capability to perform well without making any assumption about the form of the underlying distributions. To overcome the heavy computational load of the method, we recently proposed a fast approach based on a tracking mean shift estimator. In this paper we study the kernel selection problem for the mean shift background tracker. Comparative results for the Gaussian and Epanechnikov kernel are included.