Analysis of Functional Safety of an ECU implementing a Neural Network for Autonomous Driving

Today, it is impossible not to heard about Autonomous Vehicles(AVs). Autonomous Vehicles (AVs) being an essential part of the future smart city traffic is a hot topic in recent years, but still to be developed. These systems, also called self-driving cars are a kind of mobile robot and comprising perception system, behavioral decision, motion planning and intelligent control. Along with them a lot of social benefits are carried out. First of all it is expected to reduce the number of accidents, as over 90 % of accidents is caused by human mistakes. Second AVs can affect travel costs and labor costs by exploiting for example intelligent car pooling. Thirdly, they can help regulate traffic jams, that are typically related to a disrespectful interaction, adoperated by humans, of traffic signs and consequently traffic violations. Lastly, but not for importance, they will improve fuel efficiency and lower carbon emissions. Although the autonomous industry is under rapid development, AVs are still not fully replacing Human-Driven Vehicles (HVs). It is expected that the deployment of such systems is going to take only the 50% of the overall traffic by 2040. It is clear that driving safety is a need of today’s road transportation vehicles and it can be achievedthrough designing appropriate control systems with a minimumdecision-making error. A designed control system must perform indifferent road conditions and situations, and this is still challenging in the automotive industry. This thesis work is based on the analysis of the aspects of the security related with the implementation of a Neural Network for Autonomous Vehicles, and aims to: •Explain the current state-of-art regarding the latest Neural Networks training approaches for Autonomous Driving. •Deal with issues related to the safety regulations like Functional Safety and SOTIF (Safety Of The Intended Functionality). •Describe critical issues and limitations in the process of training of Neural Networks. •Suggest a solution to overcome failures and faults or limitations of the Neural network training process.