According to WHO’s report, there are over 270,000 people who are involved in traffic fatal accidents. This figure accounts for 22% of traffic fatalities in the world in 2013. To reduce those pedestrian accidents, many countries apply the pedestrian protection tests for the regulation and the third-party evaluation. For that reason, a method to design the pedestrian protection performance efficiently is required for many cars sold all over the world by automobile manufacturers. In recent years, there are some cases to assist or automate the development using the artificial intelligence acquired by the machine learning. The authors investigated whether it is possible to predict pedestrian head protection performance without using tests or CAE (Computer Aided Engineering) in this research.

The authors used the bonnet hood structures compatible with pedestrian protection and head injury value obtained from CAE for the training data. As for the hood structure, the data obtained by converting 3D geometry into a 2D image was used as the input data. Head injury value was examined by both classification and regression as output information. For the learning model, LeNet-5 of CNN (Convolutional Neural Network) was used, and the layer structure of the model was modified to be suitable for learning of pedestrian protection.

Using the learned model and validated it with some unknown hood images, the model predicted the pedestrian NCAP (New Car Assessment Program) score with an error less than 5% compared with CAE results. Also the predicted head injury criteria map agreed with accuracy more than 75%. In addition, LeNet-5 showed shorter computation time and higher accuracy when comparing the other algorithms.

Although the model was able to reasonably predict the head injury value in the center area t of hood, the accuracy of the perimeter area tended to be lower. Since the data around the perimeter area used for learning was small amount, it is considered that the accuracy is low. In future study, it is necessary to add such data or to device a method to improve accuracy even with the small amount of data.