Robust Localization for an Autonomous Racing Vehicle

In the framework of autonomous driving, one of the most challenging problems to be solved for an autonomous system which aims to navigate without any human support, is represented by the localization task. Considering the self-driving vehicle as a robotic system, the localization problem arises with motion in an unknown environment and relates to the capability of the autonomous system to know at each instant its own positioning with respect to a fixed reference. The knowledge of position and orientation of a mobile robot is referred to as pose and it is related to the state estimation capabilities of the control system of the robot. In autonomous navigation and driving, state estimation becomes even more relevant and is a key feature to achieve robustness, repeatability and high-end performances of control actions. In this scenario, autonomous racing represents the most challenging ground for testing the effectiveness of autonomous algorithms, thanks to the wide range of dynamical and physical conditions presented by the racetrack, as well as the goal of achieving the minimum lap-time. The main aim of this thesis work is the deployment of a robust and cost-effective localization architecture for the first autonomous racing prototype of Politecnico di Torino, participating to Formula Student Driverless (FSD) competition. Starting from the localization problem statement, the thesis focuses on the investigation of the main methods to achieve robust positioning and effective state estimation. The solution of the localization problem is guaranteed by the usage of well-known filtering techniques, exploiting both self-contained information and global measurements of vehicle's states. The investigated method proposes and compares the application of Extended Kalman Filter (EKF) and Unscented Kalman Filter (UKF) to the sensor fusion of on-board data streaming from an Inertial Navigation System (INS) and a Global Navigation Satellite System (GNSS) sensor. The design of the hardware layout and the software architecture are both presented, discussed and experimentally validated in real time, using a properly instrumented all-wheel drive electric racing vehicle, developed by Squadra Corse PoliTo. The proposed algorithm deployed on an on-board high-performance computing platform, has proven to achieve sub-meter accuracy in positioning tasks, also for short outages of GPS signal. Besides, the whole architecture provides precise state estimation for the retained single-track vehicle model, exploited for further control strategies. The experimental results show a substantial equivalence of the application of the two filters considered. Nevertheless, the UKF-based method is characterized by a lower estimation variance in the localization task, providing more robust results and thus is chosen for the final implementation on the vehicle.