The traditional kernel correlation filter (KCF) tracking algorithm is the best one of the discriminant tracking algorithms. However, the KCF algorithm is not robust in case of the target scale change and occlusion. Therefore, we improve the algorithm from three aspects: Improving feature expression ability, increasing scale matching strategy and anti-occlusion. First, we combine the histogram of oriented gradients (HOG) feature and the hue-saturation-value (HSV) feature. Then, a dynamic selection scale pool is employed to replace the way of fixed size matching of filters. Finally, the occlusion of the target is measured by the change rate of maximum response of the filter. And based on the target information, EdgeBoxes and the perceptual hash algorithm are used to retrieve the target and update the filter. Results of the experiments on the benthmark video trocking dataset indicate that comparing with other trackers, the proposed algorithm can effectively improve the tracking accuracy and the robustness even the target with scale change and occlusion.