A practical and reliable capability for autonomous navigation needs to reduce operation cost, to improve operational efficiency, and to increase mission safety. Celestial navigation is a very attractive autonomous navigation solution for deep space spacecraft. There are mainly two kinds of celestial navigation methods: the direct calculation method and the filter method. The accuracy of the direct calculation method is low and very sensitive to the measurement noise. The filter method can provide a better navigation performance if a high accuracy dynamical model is available. However, the main practical problem existing in the autonomous celestial navigation of spacecraft on a gravity assist trajectory is that the accuracy of trajectory model is not enough to be used in the real navigation sometimes, which may introduce large estimation error and even cause filter divergence. To solve this problem, a new celestial navigation method is proposed in this paper, which effectively combines the direct calculation method and the filter method using an interacting multiple model unscented Kalman filter (IMMUKF). The ground experimental results demonstrate that this method can provide better navigation performance and higher reliability than the traditional direct calculation method and filter method.