Learning Based Square Root Unscented Kalman Filter for Indoor Vehicle Positioning

Unscented Kalman filter (UKF) depends on system model in indoor vehicle positioning. However, the constructed model does not always agree with the actual model because of the inherent error due to lacking complete cognition for the target system. Gaussian process regression based square root UKF (GPSRUKF) can lower the error and improve estimation precision. The proposed method is based on the fact that Gaussian process regression learns the deviation between the model and the actual sampling data. Additionally, square root UKF (SRUKF) overcomes the defect that UKF requires more numerical stability for state covariance updates which need to be re-factorized to obtain the sigma points set in every iteration. Experiments on indoor vehicle positioning shows GP-SRUKF has a better performance by compensating deviation comparing with SRUKF and UKF only in nonlinear systems when we build the target model by sampling data in advance.