Aiming at the shortcoming of feature space representation in traditional mean shift, we propose an improved object tracking method. At first, the target model region is segmented into overlapped square, and their histograms are computed. Then, the feature space is constituted which has introduced spatial information into. So the accuracy is enhanced. After computing the feature space of target candidate region, the mean shift is employed to find the new target location. The result shows that the improved method can track the object more robust, accurately and quickly.