Are there similarities between the spatial patterns of autonomous and human-driven vehicle crashes? Evidence from a data-driven case study in California

Comparisons between autonomous vehicles (AVs) and human-driven vehicles (HDVs) have been widely adopted as a means of assessing AV safety. This study provides a new data-driven perspective for such comparison-based assessments by quantitatively measuring the similarity between the spatial patterns of AV and HDV crashes. Using crash records of both vehicle types as inputs, the proposed method comprises three main steps: (1) reconstruction of investigated road network for in the absence of publicly disclosed AV operational routes; (2) crash hotspot identification using Network Kernel Density Estimation (NKDE); and (3) quantitative spatial similarity analysis based on multiple similarity indicators. The method is applied to case studies in two urban areas in California, San Francisco and the Silicon Valley area, across three time periods: 2016–2018, 2019–2021, and 2022–2024. The results reveal three key findings. First, the spatial patterns of AV crashes gradually become more similar to those of HDV crashes over time. Second, AV crash spatial patterns exhibit substantially greater temporal variability than their HDV counterparts. Third, notable differences are observed between the two urban areas in terms of the spatiotemporal patterns of vehicle crashes. These findings are plausibly attributable to the expanding scale and spatial coverage of AV operations, as well as ongoing advances in human-like AV control and decision-making mechanisms over time. This study provides practical insights for urban traffic management and AV manufacturers by improving understanding of how AV crash risks evolve relative to conventional traffic systems.