A hybrid low cost approach using Extended Kalman filter and neural networks for real time positioning

Global Positioning System (GPS) and Inertial Navigation Systems (INS) are usually used for real time vehicle positioning. Unfortunately, this high cost solution still suffers from GPS outages due to multipath errors. To perform data fusion, the Extended Kalman filter (EKF) has been widely adopted as an optimal estimation tool for non-linear systems. However, it has some drawbacks in terms of stability, immunity to noise effects and observability. Furthermore, the filter is unfluenced by the vehicle dynamic variations and environment changes. On the other hand, neural networks can map input data to output data without any prior knowledge of the relationship involved. This paper presents a hybrid EKF and neural networks approach to improve the real time vehicle positioning performance while using GPS with nothing else than an odometer and a gyrometer low cost aiding sensors. While GPS signals are available, the neural networks are trained on different dynamics and outage times to learn the position errors so they can correct the future EKF predictions during GPS signal outages. The simulation results of this approach show a minimal improvement of 39% over the EKF predictions.