Worst-case scenario communications and algorithms for Collision Avoidance in Autonomous Vehicles

The world of autonomous driving is riddled with numerous challenges accompanied by a dire need of hardware and software solutions. The ultimate goal in this dissertation is to find the most efficient and effective ways to avoid collisions, manage worst-case scenarios and determine whether more research and development of new technologies is needed or if the existing systems already provide enough accuracy and reliability to be safely applied to our roads and infrastructure. Real-time processing and communications issues such as communication latency and especially processing time are thus analyzed to ensure real-time decision-making by the autonomous driving systems: on the physical layer, protocols used to reduce packet collisions and exchange messages between vehicles and other connected devices are explained in detail, without forgetting the aspects of collaboration and security. Real world driving conditions are often unsafe and unpredictable so exploring the main predictability factors is of fundamental importance to enable vehicles to detect dangerous situations ahead of time and thus process the collected information to deploy appropriate countermeasures: technologies used to share such data and enable multi-vehicle networks are discussed and compared. On-board emergency systems’ efficiency on avoiding collisions or aiding evasive manoeuvres is proven by when and how such systems intervene, so in this thesis we are going to be taking a look at what happens when the need for decisions in critical moments arise, a problem at the heart of autonomous driving. Suites of sensors are then scrutinized in both hardware and software capabilities by conducting tests on various environments where we will prove which type of sensors are appropriate for autonomous driving and what characteristics are needed to challenge changes in temperature, lighting and humidity, where particular attention to object detection accuracy is needed. Data collected by cameras and sensors also need to be compressed: we will explain why that is, what methods are the most commonly used and which ones are actually the most effective for our purpose. By comparing performance metrics between the most used algorithms, the technique proven to be the most fit- ting compression method for cloud-computing platforms and V2X communications will be determined together with mathematical methodologies used to detect and recognize objects during the drive cycle. A lot of tests run on real-world driving conditions have also highlighted how control systems perform differently based on their design principles, so the most suitable vehicle’s model and neural network learning algorithm need to be chosen and utilized based on scenarios and challenges regarding interaction with road users. Through these analyses we will prove how the solution to hardware integration is cloud processing through vehicle- to-everything communications: aspects such as data handling, rate of loss and compression, but also answers to issues regarding security, reliability and legal responsibility for data leaks will be discussed. The current available approaches to wireless communications will be explored in detail, focusing on the most important aspects such as data synchronization, packets collision, retransmission, confidentiality and authentication to ensure maximum reliability while adopting next generation hardware for networks based on 5G and 6G radio waves signals. Advancements on energy efficiency and network switching between different driving environments is debated, as well as how to solve the vulnerable road users’ low range collision prevention problem through human-machine interfaces. [...]