Optimal color space based probabilistic foreground detector for video surveillance systems

Abstract

Foreground detection is one of the well and widely studied research topic in the field of computer vision. However, it still fails to cope with the many practical issues such as illumination changes, dynamic backgrounds, and shadow. This paper proposes optimal color space based probabilistic foreground detector. The intuition is to employ two most widely used color spaces (RGB and YCbCr) one at a time to model background. A decision criteria to select optimal color space is based on mean squared error (MSE). Initial frames (say 100) without any foreground information are used to compute MSE for both color spaces. Color space with minimum MSE is selected as optimal color space (OCS). Afterwards, OCS is used to model background and detect moving information. Gaussian Mixture Models (GMM) based foreground detector is used for the purpose. Furthermore, foreground mask is cleaned from undesirable noise using morphological operations. The proposed method is tested using change detection dataset. It shows promising results and outperforms conventional GMM.