Deep Learning for Social-Aware Autonomous Vehicle-Pedestrian Interaction

Abstract

In recent years, Autonomous Driving technology has surged in popularity, becoming a key research focus. Despite commercial advancements, many legal and technical challenges remain, particularly in Avs’ interactions with human road users.

Motion control algorithms for AVs in pedestrian scenarios are crucial for safety and reliability. Traditional algorithms, which rely on manually designed policies, scale poorly with complexity and are costly. In contrast, Deep Reinforcement Learning (DRL) allows for automatic policy learning. This thesis explores automated AV decision-making in AV-pedestrian interactions using DRL and social psychology. Firstly, we propose a framework based on Social Value Orientation and Deep Reinforcement Learning capable of generating decision-making policies for the AV with different driving styles. Adding a social term in the reward function design allows us to tune the AV attitude towards the pedestrian from a more aggressive to an extremely prudent one. This model achieves a 0% collision rate and exhibits 10 behavioral modes.

We also introduce a novel pedestrian model incorporating situational awareness into a Social Force Model, enabling realistic pedestrian reactions to AV actions. We perform experiments to validate our framework and we conduct a comparative analysis of the policies obtained with two different model-free Deep Reinforcement Learning Algorithms. Comparative analysis of policies from two DRL algorithms, SAC and PPO, reveals that SAC-trained vehicles stop 30% earlier and maintain a 1.5-meter larger distance from pedestrians, while PPO policies yield 20% smoother acceleration profiles. Extending to multi-agent settings, we employ Graph Convolutional Networks to manage multiple vehicles and pedestrians, capturing agent inter-relationships efficiently.

Lastly, we develop a 3D Virtual Reality environment for studying pedestrian interactions with vehicles. Using VR technology, we collect data safely and cost-effectively. Graph neural networks predict pedestrian trajectories with a 0.17 m average displacement error, demonstrating the framework’s effectiveness in studying pedestrian-vehicle interactions.

Date of Award	24 Oct 2024
Original language	English
Awarding Institution	Northumbria University
Supervisor	Yifeng Zeng (Supervisor), Chongfeng Wei (Supervisor), Hubert Shum (Supervisor) & Edmond Ho (Supervisor)