Vehicle localization combining non-linear state observer with artificial neural network

Abstract

Abstract The developing of autonomous drive is needed to make people life more comfortable and safer, and one of the important skills to make possible the reliability of the all control system is a good localization of the vehicle. In this study, a no-linear state observer was developed using the Unscented Kalman Filter (UKF) algorithm, to estimate the global position, global orientation, and local speeds of a car inside a known path. A characterization of the sensors input measures was made and the measures of longitudinal and lateral vehicle speed were added using an Artificial Neural Network (ANN) trained in simulated manoeuvres. In this way, it was possible to reduce the error that the observer make on the estimation of the lateral vehicle speed, and so of the side slip angle, making possible an improvement of the control activity. To assess this increase in performance, a Montecarlo analysis was made comparing the architecture proposed, ANN+UKF, with state observed, UKF, with no input measure of lateral speed. The tests were done in co-simulation environment of Vi-Grade’s CarRealTime software and Matlab-Simulink.