Vehicular Dead Reckoning Based on Machine Learning and Map Matching

Global Navigation Satellite Systems (GNSS) are used today in various contexts as a source of data for several applications. They provide real-time positioning based on the transmission of electromagnetic waves from a satellite to a receiver, being subject to several factors. Some scenarios, such as canyons (urban or geographic), forests and tunnels, are challenging, since the coverage in them is unavailable or unreliable, producing rogue positioning information or no information at all. Thus, applications that demand high availability usually employ other sensors. Nevertheless, reducing the amount of such devices results in lower costs and energy consumption. Aiming to improve the reliability and availability of GNSS-based systems retaining cost-effectiveness, this work proposes a dead reckoning system, using the last known location and sensor data to infer the current position. The sensors employed here are largely available in commercial vehicles. We calculate the estimates using machine learning models and improving the results through a map matching procedure. The results, based on simulations with real GNSS and sensor data, indicate that the system is able to closely reproduce trajectories for over a minute. The obtained mean error is of approximately 19 meters, suitable for obtaining approximate locations in scenarios with unreliable satellite coverage.