A new tracking framework based on online learning was presented to solve the problem of lack of supervision in current tracking method. The proposed technique uses a tracker based on SURF, and a detector based on random forest. During the tracking, the online learning runs simultaneously from positive samples which were selected based on the similarity threshold. This mechanism abstain the decrease of detection precision from learning of negative samples. The detector can find the target and re-initialize the tracker after its failure. Then a strong tracking solution is obtained by the integration of a tracking algorithm, an online learning process and a detection algorithm. Experimental results of different video show that the proposed method can greatly improve the robustness of target tracking method under complex surrounding such as appearance change, partial and full occlusions.