Head Injury Prediction for Child Pedestrian Using Multimodal Deep Learning

Abstract

Expanding the advanced automatic collision notification system to include pedestrians is expected to help reduce the number of traffic fatalities. Accordingly, we propose a pedestrian head injury prediction method for adult males based on pedestrian collision images obtained via simulations. However, in real-world scenarios, pedestrians are not limited to adult males and range from children to the elderly. Specifically, children have short statures and the scope of body recording with the dashcam is narrowed, which renders injury prediction difficult. This study also confirms the accuracy of the proposed method for predicting head injuries in child pedestrians. The effects of including car velocity and impact position in the input information are investigated. The results show that the prediction accuracy is 88.29% when only collision images are used as inputs and that this accuracy can be improved to 92.95% by adding car velocity information.