Direct Geometrical Map to Low-Level Grid Map Registration for Robust Online Localization

In autonomous driving, robustness and precision can be increased through additional information sources in the form of prior maps. To benefit from these maps, precise localization is necessary. Localization enables the vehicle to drive along planned trajectories using map information like lane geometry and static environment information. State-of-the-art localization methods often use feature based maps or rely on dense appearance based maps for direct, association free matching. In this paper we combine both ideas. We solve the localization problem by matching dense measurement grids from LiDAR sensors directly against a geometrical, feature based map. Reflectivity and occupancy grids are obtained from multiple Li-DAR sensors. In order to avoid information reduction through feature extraction, the low-level grids are registered directly against a high-level geometrical map (Fig. 1). This new concept is solved using an exhaustive search in combination with a multi-level-resolution search. Real-time capable registration is achieved by exploiting the computational power of modern GPUs. To realize an online localizer, a Kalman Filter (KF) fuses the registration results with odometry information, obtained from serial production sensors. To ensure a fixed runtime, the registration resolution is adapted based on the state uncertainty. The robustness and accuracy of the developed methods is evaluated in extensive experiments. The localization proved to be accurate and robust against challenging situations like difficult maneuvers or missing localization features in the map. Thus, our new approach is well suited for autonomous driving applications.