In the autonomous driving environment, the attribution of responsibility becomes complex when multiple crash parties and factors are involved. This study proposes a method to attribute the responsibility of the primary crash vehicle when human drivers and autonomous driving systems coexist, and apply it to the existing Autonomous Vehicles (AVs) crash data from 2019 to 2023 in California, USA. Firstly, a causal network is constructed by integrating the Decision-making Trial and Evaluation Laboratory (DEMATEL) and Interpretative Structural Modeling (ISM) methods to analyze the mutual impact of factors in the crash data. Secondly, Random Forest (RF) is used to obtain the feature importance in AV crashes. Based on the relationship between factors and the main responsible parties, the responsibility among relative stakeholders can be quantified. Under the research data in California, in autonomous driving mode, human drivers of the primary crash vehicle and software developers both account for 31% of the crash. Following behind are other stakeholders at 21% and vehicle manufacturers at 17%. On this basis, adjustments can be made to the responsible proportion in relation to a specific crash. By identifying the impact factors of AV crashes and responsibility attribution, this study offers important insights into safe autonomous driving tests, AV production regulation, and the development of crash responsibility policies. The methodology framework developed in this paper is universal and can be applicable to AV crash analysis in diverse regions and AV penetration rates.
Wang et al. (Mon,) studied this question.