For natural human-robot interaction, we need to know location and shape of facial feature in real environment. In order to track facial feature robustly, we can use the method combining particle filter and active appearance model. However, processing speed of this method is too slow. In this paper, we propose two ideas to improve efficiency of this method. The first idea is changing the number of particles situationally. And the second idea is switching the prediction model situationally. Experimental results is presented to show that the proposed method is about three times faster than the method combining particle filter and active appearance model, whereas the performance of the proposed method is maintained.