A method for evaluating collision avoidance systems using naturalistic driving data

This paper describes a method for use in evaluating the performance of collision avoidance systems (CASs) using naturalistic driving data collected during real crashes and near-crashes. The method avoids evaluation of algorithms against specific assumptions of reaction times or response inputs. It minimizes interpretation of the involved driver's perception and response levels which permits generalizing findings beyond the performance of the involved driver. The method involves four parts: input of naturalistic crash data into alert models to determine when alerts would occur, kinematic analysis to determine when different responses would be required to avoid collision, translation of the time available into an estimate of the percentage of the population able to avoid the specific event, and an evaluation of the frequency of alerts that would be generated by the CASs. The method permits comparison of CAS performance and provides guidance for CAS development. The method is described primarily in the context of Forward Collision Warning CASs, but is applicable to other CAS types.