An Improved Background Modeling for Video Segmentation

In the existed background modeling methods, specified threshold parameters need to be given by user when the probability map is established in order to obtain the optimal segmentation results. In this paper, a new method is proposed, in which the feature intensity function of video adjacent relationship is calculated adaptively according to video features. It will avoid the drawbacks that video adjacent relationship limited with empirical value will not be adaptive to deal with various videos. Firstly, the Gaussian classifier modeling method in pixel-level is used to construct classification models for video background, video shadow and video foreground respectively. Meantime, the Ising model is used to model the adjacent relationship between pixels in video. Then the feature function of each model is calculated separately. Finally, the conditional random field model is used to bind the energies of these feature functions. The Gibbs sampling algorithm is applied to solve the model and obtain the global optimized segmentation result. Experimental results show that the parameter adaptive algorithm can approximate the segmentation results using the optimal empirical parameters.