Real-time object tracking based on sparse representation and adaptive particle drawing

Sparsity-based trackers describe a candidate region by solving the l1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{1}$$\end{document} minimization problem. This process is done for many candidates generated in a particle filter framework. This results in a high computational cost, and thus it can preclude the use of these trackers in real-time applications. To tackle this issue, we proposed a novel method to draw particles from a Gaussian distribution which not only reduces the number of candidates but also maintains the accuracy of object tracking. In the proposed method, the appearance and motion information of object is used to adaptively estimate the number of particles in each frame. In addition, a normalized correlation (NC) filter is integrated to the sparsity-based tracker in which they cooperate with each other to increase the performance of object tracking. In other words, the l1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{1}$$\end{document} minimization problem is controlled by a NC filter and the template region of correlation filter is set by a sparsity-based tracker. The occlusion degree over object tracking is accurately estimated by the sparse coefficients that its energy is adaptively controlled in the l1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l}_{1}$$\end{document} minimization problem. When an object is partially occluded, the templates of dictionary are adaptively updated based on occlusion degree which causes to increase the accuracy of object tracking. Experimental results on two tracking benchmark datasets demonstrate that the proposed method performs favorably against several popular trackers.