Abstract:

Road maps for vehicle control and navigation systems are typically generated by mapping systems that are highly accurate but updated infrequently. However, changes to the roads are made at a higher frequency. Stored road maps may therefore not capture the true road well. To resolve this, we consider online road-map estimation using the type of sensors found in production cars. The map estimation for a given vehicle is based on a global positioning system, camera, steering wheel, and wheel-speed sensors. As each vehicle covers a limited amount of road, we leverage crowdsourced map estimates from multiple vehicles to get a more complete representation of the road map. High-fidelity simulation results indicate a reduction of the estimation error of roughly 15% when using 5 agents compared to the best single agent. Furthermore, we show that the method is capable of updating map segments that have large errors, for example, as may occur during road maintenance.