Evaluation of AI methods for scenario variation to support the validation of highly automated driving functions using real measurement data

Abstract

Ensuring the safety and reliability of highly automated driving (Society of Automotive Engineers level standard (SAE) Level 3) functions is one of the most critical challenges in developing au tonomous driving technology. As these systems advance, they present complex validation challenges for obtaining series approval. This complexity arises from the need to replace a human driver, broadening the scope of scenarios that must undergo rigorous testing. Scenario-based testing solves these issues by validating these systems through the recreation of specific safety-critical driving scenarios at various levels of abstraction, from pure simulation methods(i.e., X-in-the-loop (XiL)) to real-world drives (field operational tests). The current state-of-the-art approach depends on parameterized models to generate these scenarios. While effective, this method demands extensive modeling efforts and may introduce biases in safety assessments due to the idealizations in safety relevant driving maneuvers. Moreover, current techniques do not automatically generate intuitive models for vehicle trajectories within these scenarios. In this master’s thesis, we evaluated various Artificial Intelligence (AI) methods to determine their ability to generate realistic and diverse driving scenarios, with a focus on three scenarios: cut-in, cut-out, and cut-through. Using real measurement data, we trained various generative models, including Variational Autoencoder (VAE)s, Generative Adversarial Network (GAN)s, and hybrid architectures (Autoencoder-Generative Adversarial Network (AE-GAN)). We then evaluated the models using quantitative and qualitative measures, including various metrics and visualization methods. The VAE model with a Convolutional Neural Network (CNN) was the most effective in terms of performance efficiency and computational resource use. We then benchmarked this top-performing VAE model against the traditional mathematical model within our existing testing and validation framework. The research identified the benefits and drawbacks of each AI model, providing insights into which is best suited for crafting driving scenarios that can enhance the testing and development of Advanced Driver-Assistance Systems (ADAS). The integration of the VAE model demonstrated that it could depict the characteristics of real driving maneuvers more realistically than the mathematical model used so far. Keywords: advanced driver assistance system; real traffic situation; neural networks; variational autoencoder; cut-in maneuver; cut-out maneuver; cut-through maneuver; generative adversarial networks; generative modeling; machine learning; synthetic data; software-in-the-loop simulation