Foreground/background segmentation with learned dictionary

Video sequences are viewed as a temporal collection of inverse problems. This parallel with the classical inverse problem of denoising brings us to investigate a sparse representation based approach for background subtraction. A global trained dictionary is obtained using a k-means classifier and using the matching pursuit method a set of coefficients is estimated. By linear combination of dictionary vectors (atoms) and the set of coefficients a background estimate is computed for each frame to obtain the foreground-background segmentation. The global dictionary and the coefficients are propagated and updated along the sequence. The approach yields surprisingly preliminary results, encouraging for further investigations on the possible extensions of the algorithm.