Separating background and foregroundin video based on a nonparametric Bayesian model

Separating background and foreground in video is a fundamental problem in computer vision. We present a Bayesian hierarchical model to address this challenge, and apply it to video with dynamic scenes. The model uses a nonparametric prior, a beta-bernoulli process, for both the background and foreground representation. Additionally, the model uses neighborhood information of each pixel to encourage group clustering of the foreground. A collapsed Gibbs sampler is used for efficient posterior inference. Experimental results show competitive performance of the proposed model.