Potential Crash Rate Benchmarks for Automated Vehicles

Goodall, Noah J. · 2021 · Crossref

DOI: 10.31224/osf.io/xbn6g

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the lack of consensus regarding safety benchmarks for automated vehicles (AVs), specifically determining how safe an AV must be before public deployment. The author argues that using the average human driver as a baseline is insufficient because it includes illegal and impaired behaviors like drunk and distracted driving. Consequently, an AV matching the average crash rate would statistically perform between drunk and sober driving levels. To establish more rigorous and socially acceptable targets, the study explores three perspectives: crash rates of model drivers derived from naturalistic driving data, public risk acceptance via stated preference surveys, and safety comparisons with other transportation modes. The methodology utilizes data from the Second Strategic Highway Research Program Naturalistic Driving Study (SHRP 2 NDS), which recorded over 30 million miles from 3,500 drivers. By analyzing odds ratios for risk factors such as distraction, fatigue, and impairment, the author calculates the crash rate for a "model driver"—defined as sober, rested, attentive, and cautious. Additionally, the paper reviews stated preference surveys, including Liu et al.’s study on acceptable fatality rates and Nees’ survey on driver safety percentiles. Finally, the author compares crash, injury, and fatality rates across various transportation modes (e.g., buses, commercial aviation, rail) using metrics such as vehicle-miles traveled (VMT) and person-miles traveled (PMT) to identify comparable safety baselines. The findings indicate that model drivers crash at a rate of 1,347 per 100 million vehicle-miles, representing a 33% reduction compared to the average driver’s rate of 2,020. Stated preference surveys suggest the public expects significantly higher safety standards; respondents deemed AVs acceptable only if they reduced crash rates by 80% (tolerable risk) or 99% (broadly acceptable risk) compared to human drivers. Among transportation modes, buses are identified as the most appropriate benchmark due to their similar operating environments and regulatory oversight, despite having higher crash rates per mile than passenger cars due to lower speeds and urban density. The paper presents a range of potential benchmarks in Table 4, including targets for all crashes, police-reportable crashes, and occupant/non-occupant injuries and fatalities. The significance of this work lies in providing policymakers, regulators, and developers with concrete, evidence-based safety targets. The author concludes that relying on a single metric is ill-advised and advocates for a multi-metric approach that ensures equitable safety benefits for all road users, not just vehicle occupants. Furthermore, the paper highlights the practical difficulty of proving safety via fatality rates due to the rarity of such events, suggesting that crash rates and surrogate measures are more viable for validation. The proposed benchmarks offer a framework for defining "unreasonable risk" and facilitating transparent regulation of automated driving technologies.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• incidence prevalence
• comparative international
• induced exposure
• exposure measurement
• naturalistic crash near crash
• novice drivers

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: crash risk outcomes, observational prevalence
• Methodological Resource: dataset resource