Object tracking is a hot and hard problem in the computer vision study area.We deal with large objects,which are challenged in many aspects,such as the factors of lighting, size, posture, disturbance, occlusion, and so on.The superpixel tracking method has been proposed to deal with this problem. Unlike many other approaches, it is robust in all the mentioned aspects to some extent. It is very flexible to deal with non-rigid objects just like the meanshift of color histogram does,but can be more advanced, since it takes advantage of the segmented local color histogram. Here we first introduce the adaptive superpixel tracking algorithm, which is comprised by two parts, modeling and confidence mapping using the color features of superpixels.We model them by clustering, just like the "bags of words" method does, and build the cluster confidence.The model is adaptive since it just learns from some latest tracked frames, which can accumulate errors and lead to drift easily. So we propose a refined model, which incorporates the kalman filter's ideas to this problem, by integrating the current model and the new model as an evolutionary one, to better adapt to the object variation and disturbance in subsequent frames, thus achieve more stable tracking. The evolutionary model is achieved by reclustering the cluster centers of the two models, to make new cluster centers and new cluster confidences. We allocate different weight to them, if the current model gets more weight, then the evolutionary model will be more stable, otherwise it will be more adaptive. Finally we give some experiment comparisons between the evolutionary model and the adaptive one. For most cases, when the scene of the object is stable, namely there is no big sudden light change or color change, the evolutionary model outperforms the adaptive one. The reason is that the adaptive one easily learns from other objects. But when the scene suffers big sudden change, the evolutionary model can’t quickly adapt to it and get failed, while the adaptive one may make it. In a word, the method is devised to achieve more stable tracking for stable scene applications.