It is essential for a Martian rover to know its position in order to fulfill a task. For there are some disadvantages of the rover controlled and navigated by ground station. The autonomous navigation and control of planet rover have become an important topics. Since the celestial navigation method has become one of the main trends of mobile robots navigation development, a new Autonomous navigation method for Martian rover based on celestial observation is described in this paper. This method only uses the angles between the stars and the center of Martian moons, Phobos or Deimos, which are got from an image of such objects taken by a star camera. At a given time these angles contains the information of rover's position, we can determine the position of Martian rover by the UPF (Unscented Particle Filter) method combined with the rover motion model. A simulation result demonstrated the validity and feasibility of this new method. The main factors which affect the precision, including measurement error and ephemeris error, are also analyzed.