A robust time-varying Kalman filter for a class of multi-input multi-output uncetain system is proposed. This method combines a time-varying Kalman filter with an adaptive neural network. It can overcome nonlinear uncertainty with the adaptive neural network trained by two error signals. The method can improve approaching precision, and the boundedness of the estimation error is proven by the Lyapunov theory. The proposed method is used to design state estimation of leader in the unmanned aerial vehicle(UAV) formation flight. Simulation results show that the method can estimate acceleration of leader flying with uncertain maneuvers. The follower can effectively track the leader. Thus effectiveness of the method is validated.