Background Reconstruction via Low Rank Tensor Factorization

This paper introduces a new method for background reconstruction. Background reconstruction from video sequences captured by a static camera can be regarded as a low rank factorization problem. Background is the low dimensional subspace restored from the higher dimensional visual data, and foreground is treated as sparse noise of unknown distribution. The existing algorithm could not deal with noise of unknown distribution effectively. Due to the limitation of the matrix decomposition which would lost space structure information, we process video data directly as higher order tensor based on low rank tensor factorization (LRTF). We put forward a new model of foreground by using Mixture of Gaussians (MoG) and Markov Random Field (MRF). Extensive experiments show that our method can effectively construct the background