Attitude Estimation Using Iterative Indirect Kalman With Neural Network for Inertial Sensors

Abstract

In this article, an iterative indirect Kalman filter is proposed to realize motion estimation based on inertial sensors. In the fusion of gyroscope, accelerometer and magnetometer measurements, it is vital to decrease the impact of linear acceleration (LA) and external magnetic disturbances (EMA) on the estimates. To this end, the proposed filter in this article performs first-order Gauss-Markov modeling for LA and EMA, respectively. Instead of simply adjusting the measurement noise covariance online, an iterative measurement strategy is presented to separate external disturbances based on the a posteriori estimation of the state during the measurement update. Moreover, a Long Short Term Memory network (LSTM) is designed to assist the filter in the disturbance estimation process. It matches the process noise covariance to the strength of perturbation and adapts to the disturbance noise covariance. The experimental outcomes indicate that the proposed algorithm provides better accuracy and adaptive ability than some state-of-art results.