Two new methods based on mean shift iterations for real-time tracking of non-rigid objects are proposed. The first method called MSIⅠ (Mean shift improved methodⅠ) adds a perturbation step in the conventional procedure. In this way, the mean shift more probably avoids stopping at the local plateau. The second method called MSI Ⅱextends into four mean shift procedures to prevent tracker from converging at the local maximum point. Experiments on face tracking show the advantages and limitations of the new approaches.