Fast Adaptive Robust Subspace Tracking for Online Background Subtraction

We propose a fast-adapted subspace tracking algorithm for background subtraction in video surveillance. While background scenes are modelled as a linear combination of basis images, foreground scenes are regarded as a sparse image. Every time a video frame streams in, two alternating procedures are repeatedly done: basis images are updated by a recursive least square algorithm and foreground images are extracted by solving the L1-minimization problem. In the aspect that this algorithm is basically an online algorithm fast-adapted to background change, which is very much required for real-time video surveillance, it is the most efficient among all the algorithms that are based on both low-rank condition (for background modelling) and sparsity condition (for foreground modelling).