7 Levels through the Garage: 3D Ego Localization for L4 Automatization

A basic functionality of an AD stack is the localization of the ego vehicle which is required to build a consistent online environment model using sensors and to incorporate HD map information into this environment model. Respecting the contradictory requirement of a smooth position for perception and planning versus a precise absolute position in the HD map, we propose a localization system split into “relative localization” and “absolute localization” modules. The relative localization module uses IMU and vehicle dynamics sensors to estimate the full 15-dimensional motion state of the vehicle. At the core, the system is switching between two Kalman Filters to provide a smooth and numerically stable solution for relative localization. On top of the relative localization module, the absolute localization module is using landmarks on the object or feature level to estimate a precise position in an HD map. We show how ARUCO markers and radar grid maps can be used as landmarks to estimate a precise absolute position. The system has been proven for closed loop control of L4 prototype vehicles in highway, urban and parking scenarios. Precision results, compared to a DGPS-INS ground truth in outdoor scenarios, and plots of the vehicle’s trajectory through a multi-level garage give insights into the achievable performance of the localization system.