Background subtraction with Kronecker-basis-representation based tensor sparsity and $$l_{1,1,2}$$ norm

Abstract

Background subtraction is an important and fundamental step in video analysis and is a challenging task due to the dynamic background, bad weather, complex moving behaviors and huge amount of data in real-life applications. To address this issue, we proposed a new decomposition model based on the tensor robust principal component analysis, which makes full use of the continuity of time and the correlation of space. The Kronecker-basis-representation based tensor sparsity was introduced into the proposed model to constrain the spatio-temporal correlation of the video background and to enhance the consideration of low-rank characteristics, thus effectively reducing the interference of the dynamic background. The $${l_{1,1,2}}$$ norm was used to constrain the sparseness of the spatio-temporal structure of the foreground, which strengthens the spatio-temporal continuity and tube sparsity of the video foreground and improves the accuracy of moving objects extraction. Experiments demonstrate that, in most cases, the proposed algorithm achieves superior performance in terms of F-measure scores and visual effect of separating the foreground and background of the video compared with state-of-the-art methods.