Enhancing autonomous vehicle decision-making through scenario-based traffic rule integration
File(s)
Author(s)
Tian, Hanlin
Type
Thesis
Abstract
Urban transportation is becoming increasingly complex due to urbanisation, rising traffic density, and growing safety concerns. Autonomous vehicles (AVs) are being developed to mitigate accidents, optimise traffic flow, and improve overall mobility. However, AVs still face challenges navigating complex traffic environments while adhering to region-specific traffic rules. Most existing AV systems lack mechanisms to systematically interpret and enforce traffic laws, particularly in high-risk scenarios.
This thesis presents a scenario-driven framework that integrates traffic rules into AV decision-making by integrating them into reinforcement learning (RL) reward functions. The framework first detects high-risk intersections by analysing aerial imagery, building geometry, and traffic data. It introduces a view percentage metric that links restricted visibility to higher accident rates. The framework builds on DeepLabV3+ or UNet++ and includes additional spatial and visual data to enhance risk prediction.
To enforce compliance, the framework incorporates a Retrieval-Augmented Generation (RAG) system that extracts and structures region-specific traffic regulations for AV training.
To evaluate the effectiveness of this integration, experiments were conducted using a simulation platform under high-risk intersection scenarios. Simulation results demonstrated that AVs trained with traffic rule constraints exhibited substantial safety improvements. Emergency vehicle collisions were eliminated, dropping from an average of 7.36 to 0 per 100 simulation episodes.
This thesis presents a scenario-driven framework that integrates traffic rules into AV decision-making by integrating them into reinforcement learning (RL) reward functions. The framework first detects high-risk intersections by analysing aerial imagery, building geometry, and traffic data. It introduces a view percentage metric that links restricted visibility to higher accident rates. The framework builds on DeepLabV3+ or UNet++ and includes additional spatial and visual data to enhance risk prediction.
To enforce compliance, the framework incorporates a Retrieval-Augmented Generation (RAG) system that extracts and structures region-specific traffic regulations for AV training.
To evaluate the effectiveness of this integration, experiments were conducted using a simulation platform under high-risk intersection scenarios. Simulation results demonstrated that AVs trained with traffic rule constraints exhibited substantial safety improvements. Emergency vehicle collisions were eliminated, dropping from an average of 7.36 to 0 per 100 simulation episodes.
Version
Open Access
Date Issued
2025-02-01
Date Awarded
01/11/2025
Advisor
Angeloudis, Panagiotis
Quddus, Mohammed
Publisher Department
Department of Civil and Environmental Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Master of Philosophy (MPhil)