A hybrid machine learning based low cost approach for real time vehicle position estimation in a smart city

Abstract

The Global Positioning System (GPS) enhanced with low cost Dead Reckoning (DR) sensors allows to estimate in real time a vehicle position with more accuracy while maintaining a low cost. The Extended Kalman Filter (EKF) is generally used to predict the position using the sensor’s measures and the GPS position as a helper. However, the filter performance tails off during periods of GPS failure and may quickly diverge (e.g., in tunnels or due to multipath phenomenon). In this paper, we propose a novel hybrid approach based on neural networks (NN) and autoregressive integrated moving average (ARIMA) models to circumvent the EKF limitations and improve the accuracy of vehicle position estimation. While GPS signals are available, we train NN and ARIMAmodels to learn the non-linear and linear structures in the vehicle position; therefore they can provide good predictions during GPS signal outages. We obtain empirically an improvement of up to 95% over the simple EKF predictions in case of GPS failures.