Robust object tracking based on accelerated sparse representation

Abstract

Recently tracking methods based on sparse representation have got a lot of attentions. But the huge computation in solving the L1-regularized least squares problem limits their application to real-time tracking. In this paper, we present a fast and robust tracking method based on sparse representation. By analyzing the sparsity of both representation coefficient and the representation error, a new model for sparse representation is proposed. We also design a reasonable sparseness-promoting initial value, which can produce significant increases in speed and efficiency. Finally, a new image metric called the Structural SIMilarity (SSIM) index is introduced into the process of template updating, which leads to a more perfect template updating processing. Experiments demonstrate that our new proposed method can work fast with a good robustness.