Control, perception and coordination algorithms for advanced driver-assistance system: optimized implementations and simulations

The future of urban mobility is undergoing changes with the development of intelligent cities and the increased use of autonomous vehicles. The transition to this new paradigm is a gradual process over several years or decades, but progress has been made through the implementation of smart sensors and communication infrastructure. The development of Advanced Driver-Assistance Systems (ADAS) with growing autonomy is also in progress. In this thesis, two main aspects are addressed: coordination algorithms to manage the smart city traffic flow and the perception-control pipeline of autonomous vehicles with specific emphasis on the localization and planning phases. With regards to the first aspect, several novel algorithms are proposed that exploit the new smart city capabilities to address typical problems such as Traffic Lights and Intersection Management, Parking Management, and Emergency Vehicles Management. The work proposed in this thesis is a study of the current situation in which autonomous or able-to-communicate vehicles and traditional vehicles that are not able to communicate with city infrastructure co-exist. This is a crucial aspect since mixing ADAS and traditional vehicles impacts the algorithm design. The proposed algorithms are tested in a simulated scenario in order to study unexpected behaviors since traffic flow is a complex system and some events can trigger unpredictable consequences. The results show that the proposed algorithms improve the city's livability by decreasing the waiting time at traffic lights, reducing the parking search time, and the emergency vehicle response time. In regards to the localization and planning stages, the emphasis is placed on the execution time of the algorithms, as it is a critical aspect. If the perception and control pipeline takes too long, the intended maneuver may become outdated due to changes in the environment, potentially causing safety risks. In light of this, novel implementations for localization and planning algorithms are proposed, which make extensive use of the GPU as an accelerator in order to reduce computational time. The GPU is leveraged to parallelize the algorithm and minimize memory access. Additionally, the use of different floating-point precision types is investigated to assess the impact on the results. The proposed implementations of the ORB-SLAM algorithm for the localization phase and the Frenet Path Planner algorithm for the planning phase show a consistent speedup, compared to the previously published CPU-based implementations of the algorithms.