Extreme value analysis for safety benefit estimation of adaptive cruise control (ACC)

As new automated features enter the automotive market, we need methods to assess their safety in a rapid, proactive, and iterative way. The traditional way of relying on crash statistics does not meet these needs. An alternative is to use extrapolation techniques designed to deal with rare events, such as extreme value theory (EVT). In this paper, we applied EVT to estimate the risk of collision with and without adaptive cruise control (ACC) during steady-state car following. We defined a Bayesian regression model to estimate the parameters of the Weibull distribution for block maxima (BM) of the brake threat number (BTN). We used a small, open-access dataset collected during a platooning experiment on a test track, with and without ACC. We found that ACC has extremely low probability to end up in a rear-end crash under normal car following circumstances. Although there is a expectation that ACC is generally safer than manual driving, we found that the relative risk of ACC was higher than the human control baseline in the dataset. The reason is that the manual control baseline represented a cautious driving style, which may not be typical in real traffic. Nonetheless, EVT can be used to measure the expected safety benefit of a vehicle system without requiring a large dataset. BTN was the appropriate safety metric to compare automated and manual driving mode as it accounts for specific brake behavior and performance.