Abstract: The attitude measurement by the fusing vision and inertial system can utilize the good reproducibility and stability of the visual measurement, and high output frequency and no ambient light interference problem of the inertial measurement. The robustness of the fusion system is weak as statistical properties of system noise and observation noise are not completely known and there may be abnormal measurement values. To solve the poor robustness of the fusion system, a robust quaternion cubature Kalman filter (QCKF) based on diagonalization of matrix is proposed. The influence of robust filter parameters on the measurement accuracy and robustness of the fusion system is analyzed. To improve the stability of numerical calculation, diagonalization of matrix is employed instead of Cholesky decomposition in standard CKF. The proposed method is tested and verified on the vision and inertial measurement system platform. The experimental results show that compared with the standard CKF algorithm, the proposed algorithm has higher accuracy, robustness and stability.