Abstract: In recent years, autonomous vehicles have gradually developed, and trajectory prediction has become a key technology for enabling safe and efficient decision-making in autonomous systems. Accurate trajectory prediction not only enhances the safety of autonomous vehicles, but also effectively improves traffic flow and driving efficiency. However, the future trajectory of a vehicle is jointly determined by the driver, the environment, and the vehicle current driving state. The complex driving environment and the interactions between vehicles constrain the accuracy of trajectory prediction. This paper summarizes the commonly used datasets and evaluation metrics in the field of vehicle trajectory prediction, then systematically reviews existing trajectory prediction methods, categorizing them into short-term and long-term prediction methods based on their prediction horizons. By examining the current research on various methods, the paper discusses long-term prediction methods based on deep learning, reinforcement learning, and intent recognition, as well as short-term methods based on physical models, and methods positioned between short-term and long-term prediction methods using traditional machine learning approaches. The advantages and disadvantages of these different vehicle trajectory prediction methods are compared. By synthesizing the latest developments, the paper proposes that high-level semantic features, such as traffic patterns and driving intent, play a crucial role in long-term trajectory prediction. Finally, in addressing the challenges of end-to-end trajectory prediction, it is proposed that integrating short-term and long-term prediction methods into a comprehensive trajectory prediction system will be a key direction for future research, contributing to the advancement of interpretable and real-time trajectory prediction systems in autonomous driving.

Key words: trajectory prediction, autonomous driving, deep learning, intent recognition

摘要： 自动驾驶汽车近年来逐步发展，轨迹预测是实现自动驾驶车辆安全高效决策的关键技术。准确的轨迹预测不仅能够提升自动驾驶车辆的安全性，还能有效提高交通流畅度和驾驶效率。然而车辆的未来轨迹是由驾驶员、环境以及车辆当前行驶状态共同决定的，周围复杂的驾驶环境与车辆间的交互制约着轨迹预测的准确性。总结了车辆轨迹预测领域常用的数据集和评估指标，针对现有轨迹预测方法预测时域的不同，从短时预测和长时预测两个方面对该领域进行了系统综述。通过对不同方法的研究现状进行研究，报告了长时域的基于深度学习、强化学习、基于意图识别的预测方法，短时域的基于物理模型的方法，以及预测时域位于二者之间的基于传统机器学习方法的研究现状，并对比了不同车辆轨迹预测方法的优势和劣势。通过综合最新进展，指出交通模式和驾驶意图等高级语义特征在长时轨迹预测中发挥重要作用。最后针对全过程轨迹预测的挑战，提出集成短时预测与长时预测的综合轨迹预测将成为未来研究的主要方向，有助于推动可解释、实时的轨迹预测系统在自动驾驶中的应用。

关键词: 轨迹预测, 自动驾驶, 深度学习, 意图识别