A framework for global vehicle localization using stereo images and satellite and road maps

We present a framework for global vehicle localization and 3D point cloud reconstruction that combines stereo visual odometry, satellite images, and road maps under a particle-filtering architecture. The framework focuses on the general vehicle localization scenario without the use of global positioning system for urban and rural environments and with the presence of moving objects. The main novelties of our approach are using road maps and rendering accurate top views using stereo reconstruction, and match these views with the satellite images in order to eliminate drifts and obtain accurate global localization. We show that our method is practicable by presenting experimental results on a 2 km road where mostly specific road features do not exist.