Providing a Safety Case for Artificial Neural Network Elements in Automated Driving Systems: A Review of Challenges

Artificial Neural Networks (ANNs) have rapidly gained popularity since the availability of powerful and economical hardware solutions, opening up applications in sectors such as automotive. Due to the limitations of traditional, deterministic software in the development of autonomous vehicles, the industry is increasingly applying ANNs for tasks such as sensing and understanding the environment of a vehicle equipped with an Automated Driving System (ADS) to allow it to plan and subsequently execute a safe trajectory in a traffic scenario.
There are safety challenges, however, in the deployment of ANNs for environmental perception tasks, and for motion planning and execution. A vehicle equipped with an ADS needs to be safe to gain the public’s trust in using and/or owning such vehicles. In automotive safety engineering the creation of a compelling safety case for the introduction of an ADS is common practice. This paper investigates the challenges in creating a safety case for the use of ANNs in an ADS and presents the status of research through a literature survey. It analyses the applicability of current automotive safety standards. Lastly, it reviews a few frameworks that offer advice on the development of Autonomous Systems using ANNs and evaluates their suitability for an ADS.