For the trackers using SVM, Avidan et al. [7] propose a tracking algorithm integrating SVM to discriminate the target from its background. Tian et al. [21] present a tracking system based on an ensemble of adaptively-weighted linear SVM classifiers based on their discriminative abilities. Bai and Tang et al. [22] propose an online Laplacian ranking support vector tracker (LRSVT), which incorporates the weakly labeled information to resist full occlusion and adapt to target appearance variation. Yet, there are still some limitations for these works. Firstly, most of them consider the classification problem on a single-patch level, which might lack flexibility and robustness when a drastic appearance occurs. Secondly, the features applied in these works are not unique enough.

It could negatively influence the tracking performance when a similar object exists. In this paper, we continue to explore the application of SVM classifiers in online visual tracking, where the input features are coefficients of sparse representation on a patch level. Thus, the patch-based SVMs are grouped for classifier modeling.As an elegant working model, sparse representation has recently been extensively studied and applied in pattern recognition and computer vision [23,24]. There are two basic problems [25]: the first one is to calculate the sparse solution of a linear system, while the second one refers to learning a suitable dictionary for approximation performance improvement. So far, the former one has been deeply exploring in visual tracking (e.g., [11�C13,15,17,20]).

Within the particle filtering framework, most of the works cast the tracking problem as searching the most likely sampling candidate of the target via l1 minimization. Mei and Ling [11] Drug_discovery apply sparse representation to visual tracking and deal with occlusions via positive and negative trivial templates. Wang et al. [17] propose a novel online object tracking algorithm with sparse prototypes, which adopts principal component analysis (PCA) basis vectors and trivial templates to represent the tracked target sparsely, and solve the problem by using an iterative thresholding method. Zhong et al. [20] develop a hybrid tracking method, where a sparsity-based discriminative classifier (SDC) and a sparsity-based generative model (SGM) are cascaded for target location estimation.

However, investigation of the second problem in visual tracking has just started. Liu et al. [12] develop a generative visual tracking algorithm with a static sparse dictionary of the target and dynamically online updated basis distribution model by K-selection, while a recent method proposed by Wang et al. [15] discriminates the target from the background based on the classification of the sparse coefficients with an over-complete dictionary without learning.