Methods for assessing the safety of autonomous vehicles

While the widespread adoption of autonomous vehicles (AVs) has the potential to drastically reduce the rate of traffic collisions, failure to verify their safe operation may expose the public to unacceptable risks. Due to the low frequency of traffic fatalities, verifying AV safety statistically via on-road testing is likely to be cost- and time-prohibitive, driving the need for alternate methods. This thesis examines four potential methods to assess AV safety: simulation, Failure Modes and Effects Analysis (FMEA), Fault Tree Analysis (FTA), and Systems Theoretic Process Analysis (STPA). The findings show two methods with potential: simulation based on data recorded from real environments, and quantitative FTA combined with a secondary analysis of the vehicle's machine learning algorithms. However, both approaches require significant amounts of data which may be expensive to gather. Further research into the safety of machine learning algorithms and further developments in AV simulation technology are required in order to develop more cost-effective methods for assessing AV safety.