Interactive Decision-Making Processes for Autonomous Driving Vehicle Systems.

自动驾驶车辆(AVs)与行人之间的相互作用是城市机动性和道路安全的重要方面。这项研究的目的是开发能够与行人、其他易受伤害的道路使用者(如骑自行车的人)以及其他道路车辆进行非平凡互动的自动驾驶系统。目标是开发这样的系统，以便它们可以在可能涉及高风险的情况下运行，即时间或空间受到限制，或者情况依赖于相互作用，例如在十字路口或斑马线。为了实现这一点，自动驾驶车辆系统(ADS)需要先进的预测模型，能够准确地预测参与每次交互的其他一个或多个代理的决策过程。有了这样的预测模型，我们需要为依赖交互的场景设计新的交互决策模型。这项研究旨在解决两个关键的技术问题：(1)如何将复杂的行为模型纳入决策；(2)如何纳入风险，例如来自不确定性估计的风险。此外，对于反病毒的预测模型来说，重要的是考虑到其他代理如何感知和推断反病毒在交互过程中自己的行动和意图。本质上，广告不仅必须全面了解另一方代理人的决策，还必须全面了解双方或多方之间的意图和反应的互惠模型，以确保在道路上顺利和安全地互动。将准确和复杂的行为模型纳入广告的决策过程将取决于对人类行为和其他代理人在动态环境中的行动的透彻理解。这通常是个人主义的，同时也取决于背景。要有效地捕捉和处理这些错综复杂的信息，需要开发先进的算法和机器学习技术。这些模型不仅应该能够识别不同的行为模式，还应该能够适应实时变化和意外情况。这涉及深入研究人类决策的细微差别，包括社会和环境因素，这些因素在实现安全和高效的自主操作方面发挥着至关重要的作用。特别是对于行人，本讨论将讨论身体姿势作为一种非语言交流信号的相关性，以及它对自动驾驶车辆在解释行人行为方面的效用。这项研究将利用最先进的计算机视觉方法来估计和跟踪人体姿势，作为潜在的行人意图队列。纳入和管理风险，特别是由自治系统环境中的不确定性引起的风险至关重要。不确定性可能来自各种来源，如传感器限制、不利天气条件或环境中不可预见的闭塞。有效的风险管理不仅涉及查明潜在的不确定性来源，而且还涉及量化不确定性并将其纳入决策过程。这要求开发稳健的不确定性估计技术和感知误差建模，以使ADS能够在给定的现实情况下进行训练。实际上，这将允许优先考虑安全问题。原则上，成功解决这两个关键挑战对于提高自主系统的能力和可靠性至关重要，使其能够在复杂的现实世界场景中导航，同时确保乘客和道路上互动人员的安全。

The interaction between Autonomous Vehicles (AVs) and pedestrians is a critical aspect of urban mobility and road safety. The objective of this research is to develop AVs which can nontrivially interact with pedestrians, other vulnerable road users (VRUs) such as cyclists, as well as other road vehicles. The objective is to develop such systems so that they can operate in situations that could involve high risk, i.e. where time or space is constrained, or where the scenario is dependent on interaction, such as at intersections or zebra crossings. For this to be feasible, the autonomous driving vehicle system (ADS) requires advanced predictive models that can accurately anticipate the decision-making process of the other agent or agents involved in each interaction. With such predictive models, we need to devise novel interactive decision-making models for interaction-dependent scenarios. This study will aim to resolve two key technical issues which arise here: (1) how to incorporate complex behavioural models in decision making, (2) how to incorporate risk, such as from uncertainty estimation. Furthermore, it is essential for the AV's predictive models to account for how other agents perceive and infer the AV's own actions and intentions during the interaction. In essence, the ADS must possess a comprehensive understanding of not only the other agent's decision-making but also the reciprocal modelling of intentions and responses between the two or more parties to ensure smooth and safe interactions on the road. Integrating accurate and complex behavioural models into the decision-making processes of ADS will depend on a thorough understanding of human behaviour and the actions of other agents in dynamic environments. This is often individualistic while also context dependent. Effectively capturing and processing these intricacies requires the development of advanced algorithms and machine learning techniques. These models should not only be capable of recognising diverse behavioural patterns but also adapt to real-time changes and unexpected scenarios. This involves delving into the nuances of human decision-making, including both social and environmental factors, which play a crucial role in achieving safe and efficient autonomous operation. For pedestrians specifically, this discussion will address the relevance of body posture as a non-verbal communication signal and its utility for autonomous vehicles in interpreting pedestrian actions. This study will leverage state-of-the-art computer vision methods for human pose estimation and tracking as a potential que for pedestrian intent. Incorporating and managing risk, particularly those stemming from uncertainty in the autonomous system's environment is crucial. Uncertainties can emerge from various sources, such as sensor limitations, adverse weather conditions, or unforeseen occlusions in the environment. Efficient risk management involves not only identifying potential sources of uncertainty but also quantifying and incorporating them into the decision-making process. This calls for the development of robust uncertainty estimation techniques and perception error modelling, to allow the ADS to train given realistic circumstances. This will in effect, allow for the prioritisation of safety. In principle, the successful resolution of these two pivotal challenges is crucial for advancing the capabilities and reliability of autonomous systems, enabling them to navigate complex real-world scenarios while ensuring the safety of passengers and those they interact with on roads.