Detecting Foreground in Videos via Posterior Regularized Robust Bayesian Tensor Factorization

Foreground detection is a critical step for separating the moving object from the background in video processing. Tensor factorization has been used in foreground detection due to its ability to process complex high-dimensional data, such as color images and videos. However, traditional tensor factorization often lacks the ability for uncertainty quantification. Bayesian tensor factorization can measure the uncertainty by considering the distributions of the tensor factorization model parameters. Besides, domain knowledge is commonly available and could improve the accuracy of foreground detection of the Bayesian tensor factorization model if it can be appropriately incorporated. In this work, a new Bayesian tensor factorization model, named Posterior Regularized Bayesian Robust Tensor Factorization (PR-BRTF), is proposed with incorporating characteristics of dynamic foreground, as a sparsity posterior regularization term. Furthermore, the variational Bayesian inference and L1 norm is combined for inducing sparsity with an efficient inference. The experiments in real-world case studies have shown the performance improvement of the proposed model over state-of-art methods.