Steering the Future: A Holistic Examination of Autonomous Vehicle-Road User Dynamics for Safer Roads

Abstract

The road to safer cities in the presence of autonomous vehicles (AVs) requires a deep understanding of the safety implications and intricate behavior patterns of diverse road users. Despite the ongoing research to improve the algorithms used by AVs to perceive and interact with the world, numerous questions still linger. Such issues limit the smooth integration of AVs into the urban traffic scene and remain mostly unresolved due to the lack of ground-truth evidence under the current early adoption rates. This dissertation focuses on the AV transition from theory into practice by utilizing real-world AV sensor and collision data to develop a crisp understanding of the potential benefits and pitfalls of human-machine interactions. The research questions are addressed through the lens of advanced modeling and assessment techniques. Specifically, road user trajectories are collected from AV sensor data to extract traffic conflicts. Then, an extreme value theory model is developed to predict AV collisions and investigate safety concerns and potential causes of risky interactions. To quantitatively represent the behavior and preferences of road users around AVs, the inverse reinforcement learning technique is employed. This approach maps the evasive actions of road users under different interaction scenarios and paves the way for advanced modeling in the future. Furthermore, innovative machine learning strategies are also introduced using this knowledge to enhance the risk-based AV hyperawareness, enabling them to predict conflicts in real time. Finally, spatial-temporal analyses of AV collisions are conducted to provide the necessary outlook into the future of AVs. Overall, the findings and recommendations of this research envision and contribute to a safer and more efficient transportation system enabled by the development and refinement of AV technology. In the face of challenges and setbacks, this proactive approach drives research in this field forward.