Deep Learning Highway Traffic Scenario Construction with Trajectory Generators

This paper presents a method for generating synthetic highway traffic scenarios. Scenarios are not only used in the training and testing phases of automated vehicles but also in the verification and validation procedures. However, obtaining a sufficient scenario database that captures the diversity of traffic is not straightforward. For increasing the size of the database while preserving the diversity, we propose a method that uses a multi-step approach with generative modeling techniques. The first stage of the method uses a Variational Autoencoder as trajectory generator. In the second stage, we make use of a Generative Adversarial Network that constructs inputs to multiple trajectory generators for obtaining complementary trajectories that form a scenario. The method is demonstrated on a lane-changing highway scenario with two vehicles. The generated scenarios are evaluated qualitatively by visual inspection using a vehicle-centered representation. Furthermore, the results are quantitatively evaluated by checking for collisions and requirements that were used to obtain the training data. Finally, the distribution of the generated data is investigated using dimensionality reduction. The results of the experiments show that our method generates multi-vehicle lane change scenarios that are similar to the ones that are found in the dataset.