To solve such problems as target scale change, deformation and occlusion in multi-target tracking, this paper proposes a new algorithm with good robustness.This algorithm adopts the improved SVM classifier for online learning, and the tracking problem is considered to be a structural learning problem with the maximum interval and the updating mode is improved to find the optimal weight. At the same time, the mechanism for the online updating of the classifier is adjusted, and the accumulated error is reduced to a certain extent, in which way online learning gives play to its advantages more adequately. In addition, the algorithm adopts the correlation filter to adaptively adjust the size of the tracking box, and proposes the mechanisms of occlusion processing and data association, so that the numbers of the targets that reappear after the occlusion are not exchanged.The experimental results prove that the proposed algorithm improves the tracking accuracy and can fulfill multi-target tracking tasks in complex background environments.