Abstract: In path planning for shared spaces between autonomous vehicles and pedestrians, accurate and efficient pedestrian trajectory prediction is critical for ensuring road safety. Pedestrian trajectory prediction not only relies on historical behavior data but also requires a comprehensive consideration of the complex dynamic interactions between pedestrians and vehicles, traffic infrastructure, and multi-directional vehicles. Significant advancements have been made in this field in recent years, making it a focal point of research. This paper provides a systematic review of the current research. Firstly, it defines the core concepts of pedestrian trajectory prediction and conducts an in-depth analysis of the main prediction methods. It then comprehensively outlines the primary data sources for pedestrian behavior, including LiDAR, cameras, and other multimodal sensing devices, while exploring key feature extraction methods, such as pedestrian motion features, contextual scene characteristics, the impact of traffic infrastructure, etc. Based on these data, this paper systematically reviews both physics-based and data-driven prediction approaches, with a focus on the development of statistical models, deep learning, and reinforcement learning models. Special emphasis is placed on deep learning methods, categorized by network architecture into sequential models, convolutional neural networks, graph convolutional networks,  generative adversarial networks, etc. This paper also reviews commonly used datasets and evaluation metrics in the field, providing a thorough evaluation of current algorithmic performance. Finally, it addresses the challenges in pedestrian trajectory prediction for autonomous driving, particularly the dynamic coupling between pedestrians with multi-directional traffic and infrastructure, offering potential solutions and discussing future research directions.

Key words: autonomous driving, pedestrian trajectory prediction, deep learning

摘要： 在自动驾驶汽车与行人共享空间的路径规划中，精准、高效的行人轨迹预测是保障道路安全的核心问题。行人轨迹预测不仅依赖于历史行为数据，更需全面考虑行人与车辆、交通设施及多方向车辆间的复杂动态交互。近年来，该领域取得了显著进展，逐渐成为研究热点。系统梳理了现有的研究成果，界定了行人轨迹预测的基本概念，并对主流预测方法进行了深入剖析。归纳了行人行为数据的主要来源，包括激光雷达、摄像头等多模态感知设备，并探讨了关键特征提取方式，涵盖行人运动特征、场景上下文特征及交通设施影响等。基于这些数据，对物理模型与数据驱动的预测方法进行了系统总结，重点分析了统计模型、深度学习与强化学习模型的发展现状，尤其是深度学习方法，依据其网络结构进一步细分为序列模型、卷积神经网络、图卷积神经网络和生成对抗网络等类型。总结了该领域常用的数据集和评价指标，对现有算法的性能进行了综合评估。针对行人轨迹预测在自动驾驶中的挑战，尤其是行人与多方向车辆及交通设施之间的动态耦合问题，提出了潜在的解决思路，并展望了未来的研究方向。

关键词: 自动驾驶, 行人轨迹预测, 深度学习