Authors: Wenbo Sun, Jingwen Hu, Carol Flannagan, and Patrick Bowman—University of Michigan Transportation Research Institute; Iskander Farooq and Anil Kalra—Ford Motor Company
Abstract
Computer simulations have been widely used for occupant and pedestrian injury prediction as a result of vehicle impacts; however, their validity is often limited by the testing data being validated against. This article describes a machine learning method to improve prediction accuracy by calibrating the simulated pedestrian injury risks with field crash injury data. The concept is to construct a surrogate model of the kinematic-based simulation model to generate fast predictions in the full input space and then search for the optimal simulation parameters and model discrepancy that match the predicted injury risks from the surrogate model with the injury outcomes in the field data. Consequently, the calibrated surrogate model integrates the field data with the kinematic-based simulation model and field data, returning more accurate predictions throughout the input space. The effectiveness of the proposed method is demonstrated by a case study where the MADYMO simulations were calibrated by the pedestrian injury data from Pedestrian Crash Data Study (PCDS).
Pages: 4
Event: 67th Stapp Car Crash Conference
Type: Short Communication