Collaborative Methods for Real-Time Localization in Urban Centers

Abstract

This article presents an effective solution for the localization of a vehicle in dense urban areas where GNSS-based methods fail because of poor satellite visibility. It advocates the use of a visual-based method processing georeferenced landmarks obtained after a learning path and stored in a new layer of the geographical information system (GIS) used for navigation. Real-time localization gives, with few failures, accurate results in the areas covered by the GIS. The integrity of the localization is obtained by running another algorithm in parallel, processing odometric data combined with the geometric model of the drivable area and, when available, GNSS data in tight coupling. An ellipsoidal confidence domain is updated by using both extended Kalman filtering (EKF) and set-membership estimation. Although less accurate, this estimation is reliable and, when the visual method fails, the availability of a confidence domain enables us to speed up the restart of the visual method while navigating cautiously. A large-scale experiment (>4 km) was conducted in the centre of Paris. We compare the absolute localization results with the ground truth obtained by combining RTK-GPS and a high-end inertial measurement unit (IMU).