High positioning precision can be obtained by autonomous navigation method based on indirectly sensing horizon by stellar atmospheric refraction. Because of the limited resolution of the atmosphere density model, large instantaneous error would occur in the horizon sensed by this method. In this case, the performance of the navigation filter would be degraded, even divergent. A fuzzy inference system was applied to autonomous celestial navigation and a fuzzy adaptive algorithm based on the Unscented Kalman filter was presented to resolve the problem above. Stronger adaptive ability of navigation filter can be obtained by this adaptive algorithm when the observation is abnormal. The computer simulation results demonstrate the validity of this method.