A statistical learning approach for estimating the reliability of crash severity predictions

Ahead of an unavoidable collision, the actual crash constellation and along with it, the crash severity can significantly change based on the driver actions. To justify the use of safety measures like airbags, prior to an accident, the severity of the predicted crash must be high enough and the crash severity prediction itself must be reliable. In this work, a machine learning driven reliability estimator for crash severity predictions is presented. The reliability estimate is obtained by simulating various driver hypotheses and analyzing the corresponding crash severity distribution. A simulation framework is introduced, utilizing a two-track dynamics model and a mass-spring model, to simulate the pre-, in- and post-crash phases and automatically generate a large amount of crash data. The data are used to train a Random Forest regression model, capable of estimating the reliability of one crash severity prediction in real-time and thereby, around 105 times faster than with simulations, and with a correlation coefficient of true and predicted reliability values of 0:92.