An approach to automate accident scenario generation using recurrent neural networks

Abstract: There is a need to improve the test procedure of Active Safety Systems through the automation of scenario generation, especially accident scenarios that are critical for testing. The purpose of this thesis is to provide an approach to automate the test generation process using machine learning. We use a recurrent neural network, applied in other domains to related problems where temporal data needs to be modelled for the generation of accident scenarios. We build a dataset of accident scenarios that occur at an intersection in a road traffic simulator and use it to train our model. We deliver an approach by testing different model parameters and input features and show generated accident scenarios in comparison to ground truth scenarios. We evaluate the quality of our generated accident scenarios through a set of metrics which we introduce in the paper.