Traget tracking methods based on the sparse representation mostly apply a template update strategy based on the fixed threshold which is difficult to adapt to the changing shape of target. In addition, sparse representation is inadequate in capturing the manifold structures hidden in target samples. A template update strategy based on the multi-level hierarchical dictionary is proposed according to drawbacks of the template update strategy based on the fixed threshold. A tracking method combining multi-level sparse representation and metric learning is proposed in order to improve the ability to distinguish between background and targets. Experimental results show that the proposed method can improve the tracking accuracy and robustness effectively.