Pedestrian Injury Case Reconstruction through Data Fusion and Machine Learning

The proportion of pedestrian injuries in motor-vehicle-crash-induced injuries in the U.S. has been increasing in recent years. Although extensive police-reported data on pedestrian injuries is available, the incomplete nature of the crash and injury information in these datasets presents a significant challenge for statistical injury analysis and pedestrian protection research. This study aims to address this issue by combining simulation data and field data to impute critical missing crash information in pedestrian crash cases through machine learning techniques. A total of 9,000 MADYMO simulations were generated using maximal projection design, incorporating variables such as pedestrian demographics, crash conditions, and vehicle impact parameters. Gaussian process (GP) surrogate models were trained to predict injury risks with simulation parameters calibrated using the complete crash information in the Pedestrian Crash Data Study (PCDS) dataset. Maximum likelihood estimations were then employed to impute the missing vehicle speed in Linked Michigan Trauma dataset. Validation involved comparing the imputed vehicle speed distribution with that of the PCDS dataset and verifying four CIREN cases reconstructed by both the proposed method and a physics-based approach. The histogram of the reconstructed vehicle speeds in Linked Michigan Trauma dataset highly correlated with that from the PCDS dataset. In the four CIREN cases, the absolute deviation between the reconstruction vehicle speeds from the proposed method and physics-based approach was 9 kph on average, with the predicted injury risks matched the observed AIS levels. These results support the use of machine learning for reconstructing missing crash data and enhancing pedestrian injury risk modeling.