随着自动驾驶的快速发展，对自动驾驶的安全性的研究也逐渐被重视，而预测其他交通参与者的未来运动是保证自动驾驶安全性的关键任务之一。然而，准确预测目标车辆的未来轨迹是一项十分困难的挑战，因为目标车辆的行为同时受到自身意图、周围智能体和交通环境的影响，而且车辆的行为还具有随机性和多模态性。因此，轨迹预测网络不仅需要对交通环境中的多智能体交互进行建模，并且需要能够预测出多模态的未来轨迹。近年来基于深度学习的方法取得了巨大的进步，然而现有方法在多模态数据建模和多模态轨迹生成方面都还存在许多问题。本文针对现有方法中存在的问题,对交通场景下的车辆轨迹预测方法开展研究。首先，针对多模态输入数据建模问题，本文提出了基于路径注意力网络的轨迹预测方法。该方法建立了一种特征的选择和融合机制，根据车道连通性等先验知识自动为目标车辆规划未来可能的路径，并使用基于路径的注意力网络建模智能体之间的交互，该方法在进行多模态特征融合时可以根据路径自适应地选择和抛弃部分特征向量，解决了现有方法中存在的冗余特征干扰问题，提升了预测准确率。其次，针对多模态轨迹生成的问题，本文在基于路径注意力网络的轨迹预测方法的基础上，引入了自监督学习的方法。本文从高精地图和路径本身蕴含的多模态性出发，构造路径聚类与分类辅助任务对特征提取网络进行预训练，使特征提取网络学习到这种多模态特征，从而增强的轨迹预测的多模态性。此外，本文还研究了该方法在复杂场景下进行多目标轨迹预测的应用潜力。通过对路径规划、智能体间注意力以及特征融合等模块的改进，实现了该方法在智能体数量大的复杂场景下的迁移应用。本文进行了广泛的实验来验证方法。在公开数据集上的实验结果证明了本文方法的有效性，其中基于路径注意力网络并结合路径聚类辅助任务的方法，在minFDE6指标上超过了所有现有方法。本文还通过消融实验和扩展实验分别检验了本文方法中不同模块的作用，以及自监督学习对模型的提升效果。此外，在复杂场景数据集上进行的实验证明了该方法具有高度的并行化和高效率。

With the development of autonomous driving, research on the safety of autono-mous driving is gradually being valued. Predicting the future movements of other traf-fic participants is one of the key tasks to ensure the safety of autonomous driving. However, accurately predicting the future trajectory of the target agent is a very diffi-cult challenge, because the behavior of the target agent is influenced by its own inten-tions, surrounding agents, and the traffic environment together and the behavior of the agent also has randomness and multimodality. Therefore, trajectory prediction net-works not only need to model multi-agent interactions in traffic environments, but also need to predict multimodal future trajectories. In recent years, the method based on deep learning has made great progress. However, there are still many problems in multimodal data modeling and multimodal trajectory generation. This paper focuses on the problems existing in existing methods and conducts research on traffic scene trajectory prediction methods. Firstly, to solve the problem of multimodal data modeling, this paper proposes a trajectory prediction method based on path attention network. This method automati-cally plans possible paths for the target agent based on prior knowledge such as lane connectivity, and models the interaction between agents using path attention networks. This method establishes a feature selection and fusion mechanism, where the network adaptively selects and discards some features based on the path, solving the problem of redundant feature interference in existing methods.Secondly, to solve the problem of multimodal trajectory generation, this paper introduces the self-supervised learning method in trajectory prediction. Starting from the multimodality inherent in high-precision maps and paths, this paper constructs pretext tasks to enable feature extraction networks to learn this multimodal feature, thereby enhancing the multimodality of trajectory prediction. In addition, this paper also investigates the potential application of this method for multi target trajectory prediction in complex scenarios. By improving modules such as path planning, agent attention, and feature fusion, this method has been applied for migration in complex scenarios with a large number of agents. This paper conducted extensive experiments to validate the method. The experi-mental results on public datasets demonstrate the effectiveness of our method, where the method based on path attention networks combined with path clustering tasks out-performs all existing methods in terms of minFDE6 metric. This paper also tested the role of different modules in this method and the improvement effect of self -supervised learning on the model through ablation experiments and expansion ex-periments. In addition, experiments conducted on complex scene datasets have demonstrated that this method has high parallelism and efficiency.