Emerging autonomous vehicle risks: The role of telematics and machine learning based risk assessment

The evolution towards fully autonomous vehicles (AVs) is set to considerably reduce road accident rates, reduce greenhouse gas emissions and liberate time spent driving. The transition is also expected to significantly disrupt the risk and liability landscape as humans are disconnected from the driving task. New, unfamiliar and unquantified risks will emerge and the increasing levels of automation will likely see a redistribution of liability. Consequently, insurance companies and regulatory bodies often lag behind in the identification, analysis, response and management of emerging risk structures. Traditional risk models are reactive by nature and inhibited by the lack of historical data relating to the likelihood or consequences of automated vehicle accidents or adverse events. Moreover, the risk structure of automated vehicles will continue to evolve, demanding proactive and adaptable risk assessment methodologies. This Thesis contributes novel, proactive methodologies to address these limitations and overcome the inadequacies of conventional, reactive risk assessment approaches. In particular, we posit that telematics data are especially suited to address these problems and demonstrate the efficacy of telematics-based risk assessment methodologies utilising several machine learning models to process the sensor generated data for proactive risk assessment. Chapters 2 and 3 comprehensively review the risk structure of semi-autonomous vehicles (SAV) and propose novel methodological approaches to processing vehicle telematics data and quantifying SAV risks for general risk assessment and risk pricing applications. These chapters focus primarily on semiautomation risk assessment and split risk structures. Namely, as control alternates between human to autonomous system, risk will migrate between technological and human related vulnerabilities. Novel machine learning methodologies are presented to extract telematics-based risk factors and generate risk scores or model frequency and severity distributions. Chapter 4 postulates a unique machine learning-based risk assessment methodology to quantify the risk exposure of semi and fully AVs relative to human benchmarks. Using telematics data, this approach allows practitioners and academics to quantify AV risk against a particular risk group of human drivers proactively. Finally, Chapter 5 proposes a unique risk management methodology that model “behavioural hotspots” using telematics data gathered from 46 study participants in Ireland and geostatistical machine learning tools. The result are used to create a novel risk-aware path planning algorithm for autonomous vehicles. The proposed methodologies and resulting applications contribute to field of automated vehicle risk assessment. They offer a proactive alternative to AV risk assessment for insurance companies, manufacturers and regulators, tasked with quantifying and ensuring the safety of automated vehicles. Each chapter in this Thesis represents a peer-reviewed journal article with a minimum and maximum impact factor of 2.5 and 5.7 respectively.