Comparability of driving automation crash databases

Abstract

Introduction: This paper reviewed current driving automation (DA) and baseline human-driven crash databases and evaluated their comparability. Method: Five sources of DA crash data and three sources of human-driven crash data were reviewed for consistency of inclusion criteria, scope of coverage, and potential sources of bias. Alternative methods to determine vehicle automation capability using vehicle identification number (VIN) from state-maintained crash records were also explored. Conclusions: Evaluated data sets used incompatible or nonstandard minimum crash severity thresholds, complicating crash rate comparisons. The most widely-used standard was “police-reportable crash,” which itself has different reporting thresholds among jurisdictions. Although low- and no-damage crashes occur at greater frequencies and have more statistical power, they were not consistently reported for automated vehicles. Crash data collection can be improved through collection of driving automation exposure data, widespread collection of crash data form electronic data recorders, and standardization of crash definitions. Practical applications: Researchers and DA developers may use this analysis to conduct more thorough and accurate evaluations of driving automation crash rates. Lawmakers and regulators may use these findings as evidence to enhance data collection efforts, both internally and via new rules regarding electronic data recorders.