Scenario-based validation is a promising approach for the safety validation of highly automated driving systems. By modeling relevant driving scenarios, utilizing simulations and selecting insightful test cases, the testing effort is reduced. However, current methods can't automatically create intuitive models of the vehicle trajectories in a scenario. We propose to use unsupervised machine learning to train neural networks to solve the modeling problem. The models learn a small set of intuitive parameters without the need for labeled data and use them to generate new realistic trajectories. The neural networks, which base on the InfoGAN and beta-VAE architectures, are adapted from the image domain to the time series domain. Although our methods are generally applicable, our experiments focus on lane change maneuvers on highways. To train the networks, we use more than 5600 measured lane change trajectories extracted from the highD dataset. Our results show that the networks learn to describe lane change maneuvers by up to four intuitive parameters. Furthermore, the networks are able to map existing lane change trajectories to values of the learned parameters and generate new, previously unseen, realistic trajectories. We compare both architectures among themselves and to a polynomial model, and show respective advantages.
Data-Driven Maneuver Modeling using Generative Adversarial Networks and Variational Autoencoders for Safety Validation of Highly Automated Vehicles
01.11.2018
3151639 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Maneuver Detection of Space Objects using Generative Adversarial Networks
British Library Conference Proceedings | 2018
|