Abstract: A new high-precision star sensor/gyro attitude determination algorithm based on nonlinear predictive filter (NPF) and second-order divided difference filter was presented. In this algorithm, model error was estimated by NPF, and the state was estimated via second-order divided difference filter by using compensated system model. While it is difficult to obtain the accurate dynamic model of attitude determination system, extended Kalman filter (EKF) treated uncertain model error as zero-mean white noise might cause low precision of state estimation even to divergence. The proposed method solved such problem. The nonlinear model was the second-order approximation while using second-order divided difference filter without calculating the partial derivatives of system model, and then attitude estimation with high precision was obtained. Simulation results show that the proposed method estimates and compensates the model error in real-time effectively and improves the estimation precision, also the filter period influences the precision slightly. The effectiveness and the robustness of the proposed method are proved.