The stationary on-line calibration Kalman filter mathematical model of dual-axis rotary inertial navigation system (INS) whose observation is velocity errors is built. The state vector includes ground velocity errors, misalignment angles and the null bias, scale errors and installation errors of inertial instruments. The observability of the calibration system in different angular motion statuses is analyzed using piece-wise constant system (PWCS) method, which concludes that the method of dual-axis rotation can improve the observabilty of the calibration filter. The state vector is reduced according to the observabilty degree analysis of singular value decomposition (SVD) method and the application characteristics of the rotary INS, and the reduce model is 12-devision. The order number reduced model decreases 55%, which can notably cut down the calculation of Kalman filter and improve the efficiency and instantaneity of the navigation computer. Simulation indicates that the estimation precision is not lower than the original model and the filtering convergence of some state are accelerated.