To overcome the shortcomings of the traditional Mean-shift algorithm for object tracking such as the localization error caused by background pixels and the tracking failure from the object occlusion,an improved Mean-shift algorithm is proposed. Firstly, according to the difference of color distribution between the object and the background in the initial frame, a log-likelihood image is set up to select the discriminative color features for object modeling, and then the candidate modeling is established by the same way. By above operation,the effect of background pixels on the image has reduced greatly.Secondly, the whole candidate region is separated into several overlapped fragments, and every fragment is iterated by the Mean-shift.Then, the object localization is reset by the location of fragment in the candidate region, which matches mostly to the corresponding fragment in the object region. Experimental results show that the fragment based on the Mean-shift is very robust to partial occlusion. Furthermore,when object is severely occluded, the linear prediction can be used to estimate the probable location of the object in the next frame. These results prove that the tracking using the improved Mean-shift algorithms has good localization precision and is robust to partial and severe occlusions.