Normalizing crash risk of partially automated vehicles under sparse data

Abstract

The safety of increasingly automated vehicles is of great concern to regulators, yet crash rates are generally reported by manufacturers using proprietary metrics with limited source data. Without consistent definitions of crashes and exposure, automated vehicle crash rates cannot be meaningfully compared with baseline datasets. The objective of this study was to establish methods to normalize automated vehicle crash rates using one manufacturer’s crash reports as a case study. The manufacturer’s quarterly crash rates for vehicles using SAE Level 1 and Level 2 automation were compared. Road type was controlled for using data from a naturalistic driving study with the same model vehicles, while driver age was controlled for using demographic ownership surveys. Although Level 2 vehicles were claimed to have a 43% lower crash rate than Level 1 vehicles, their improvement was only 10% after controlling for different rates of freeway driving. Direct comparison with general public driving was impossible due to unclear crash severity thresholds in the manufacturer’s reports, but analysis showed that controlling for driver age would increase reported crash rates by 11%. These results establish the need for detailed crash data, crash definitions, and exposure and demographic data in order to accurately assess automated vehicle safety.

Keywords:

• autonomous vehicle
• vehicle automation
• crash rate
• safety

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data used in this analysis are hosted at the Open Science Framework repository and available for download https://doi.org/10.17605/OSF.IO/5HKMD.

Notes

1 Tesla’s most recent report as of this writing, for April through June of 2021, uses different safety metrics than previous reports (Kane, 2021). Previous reports provided crash rates driving with “Autopilot engaged,” “without Autopilot but with our active safety features,” and “without Autopilot and without our active safety features” (Tesla, Inc., 2021) The latest report uses different terminology, distinguishing between driving “in which drivers were using Autopilot technology (Autosteer and active safety features)” and “drivers who were not using Autopilot technology (no Autosteer and active safety features)” (Tesla, Inc., 2021) It is not clear from their wording if how the prior category of active safety features only maps to the new categories. Without further guidance, these new crash rates cannot be compared to previously reported crash rates.

2 These figures were reconstructed from (Gershon et al., 2021) as follows. There were 22,108 total miles, of which 62% (13,707 miles) were on freeways (“limited access highways”). Of freeway miles, 40% (5,483 miles) used Super Cruise, 10% ( 1,371 miles) used adaptive cruise control, and 50% (6,853 miles) were under manual control. The non-freeways miles were calculated by subtracting freeway miles from total miles, with total mileage reported in the paper as 5,514 miles using Super Cruise, 1,891 miles using adaptive cruise control, and 14,702 miles under manual control.