aalto1 untyped-item.component.html
Data-driven diffusion models for enhancing safety in autonomous vehicle traffic simulations
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu |
Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.
Authors
Date
Department
Major/Subject
Mcode
ELEC3025
Language
en
Pages
52 + 1
Series
Abstract
Traffic simulations are essential to the development of autonomous vehicles because they provide a safe and low-cost environment to train and evaluate autonomous agents. Without traffic simulation, the development of cutting-edge autonomous driving technologies can be hindered by the fear of real-world safety consequences, such as collision with properties and pedestrians. Besides, the cost of traffic simulations is considerably lower than real-world evaluation, as gasoline and labor expenses are reduced. Within traffic simulations, scenarios that are safety-critical are the most important ones for improving safety in autonomous vehicles. The safety of autonomous vehicles is defined by the ability to navigate under near-collision (safety-critical) scenarios, however, these scenarios are rare in the real world, leading to data deficiency problems. Besides, as autonomous driving agents become better, these safety-critical scenarios happen less frequently, posing more difficulties in training autonomous driving agents against these important scenarios. In addition, high realism in simulations is crucial, because it means better translation from simulation performance to real-world performance. Recent research on safety-critical traffic scenario generation indicates that diffusion-based approaches have achieved the most effective and realistic results. However, existing diffusion-based methods overlook driver behavior complexity and traffic density information, even though traffic density information and the unpredictability of other human drivers have a great effect on driver decision-making. Therefore, this thesis aims to develop novel adversarial guidance functions for diffusion models incorporating behavior complexity and traffic density to generate more effective and realistic safety-critical traffic scenarios. The proposed method is evaluated on two evaluation metrics: effectiveness and realism. Using these evaluation metrics, the experimental results show that the proposed method has better efficacy when compared to other state-of-the-art methods.