Sparse coding and Gaussian modeling of coefficients average for background subtraction

A sparse coding based approach for background subtraction is proposed in this paper. The background model is composed from a K-SVD dictionary and a set of mean coefficients associated to each image location. Due to the use of sparse coding, our approach has a regional character. The recovered value of a pixel is obtained by reconstructing the surrounding image patch. In order to avoid problems introduced by difficult situations like dynamic backgrounds, an additional Gaussian model on the average of the coefficients set is used. A foreground confidence image results from this modeling. Two threshold will output the final background-foreground binary map. A first threshold on the confidence image selects possible foreground candidates. For these candidates we consider the reconstruction error, represented by the absolute difference between the reconstructed frame and the estimated background. A second threshold on these candidates offers the final discrimination. Our approach is tested against state-of-the-art methods. It is proved to perform better both in terms of visual comparison and quantitative measures.