Extracting Scenarios from Autonomous-Vehicle-Involved Crashes.

The burgeoning interest in autonomous driving vehicles, notwithstanding their transformative potential in the realm of transportation, engenders profound safety apprehensions owing to the potentially catastrophic consequences of mishaps. In this research, we employ a Biterm Topic Model in tandem with natural language processing techniques to scrutinize collision reports involving these technologically advanced vehicles. Upon the curation and pre-processing of a dataset comprising such reports, we implemented the model to discern recurring themes and their interconnectivity. This analysis revealed that sensor failures, software malfunctions, and pedestrian accidents emerged as the most prevalent issues. Additionally, we detected a potent correlation between certain themes and discerned spatiotemporal fluctuations in regulatory focus across diverse autonomous vehicle vendors. The implications of these findings are far-reaching for ameliorating the safety and dependability of autonomous vehicles and underscore the imperative for continual safety research in this rapidly evolving field.