欢迎访问中国科学院大学学报,今天是

中国科学院大学学报 ›› 2026, Vol. 43 ›› Issue (6): 834-846.DOI: 10.7523/j.ucas.2024.066

• 电子信息与计算机科学 • 上一篇    下一篇

基于场景上下文信息的长时多车轨迹预测

杨秋宇1, 缐凯2, 郭继孚2, 朱重远2, 焦建彬3()   

  1. 1.中国科学院大学集成电路学院,北京 100029
    2.北京交通发展研究院,北京 100073
    3.中国科学院大学电子电气与通信工程学院,北京 101408
  • 收稿日期:2024-01-29 修回日期:2024-06-25 发布日期:2024-07-17
  • 通讯作者: 焦建彬
  • 基金资助:
    中国科学院战略性先导科技专项(XDA27000000)

Long-term multi-vehicle trajectory prediction with scene contextual information

Qiuyu YANG1, Kai XIAN2, Jifu GUO2, Zhongyuan ZHU2, Jianbin JIAO3()   

  1. 1.School of Integrated Circuits,University of Chinese Academy of Sciences,Beijing 100029,China
    2.Beijing Transport Institute,Beijing 100073,China
    3.School of Electronic,Electrical and Communication Engineering,University of Chinese Academy of Sciences,Beijing 101408,China
  • Received:2024-01-29 Revised:2024-06-25 Published:2024-07-17
  • Contact: Jianbin JIAO

摘要:

精准感知周围车辆的未来行动对自动驾驶的安全性保障有着至关重要的作用。已有的轨迹预测方法可以大致分为联合预测和边际预测2类,尽管联合预测方法在多车问题上能取得更好的场景一致性,但在长时间预测上这2类方法都存在局限,因为它们无法模拟驾驶员决策随时间的变化。针对此问题,提出全新的联合预测方法(TPP),其核心是轨迹后处理模块。该模块利用注意力机制构建不同车辆间的交互,通过单车多编码的设计模拟车辆行驶中决策的更新,最终利用解码器生成具有场景一致性的多车轨迹。分别在短时预测数据集和长时预测数据集上对该方法进行评估,并与主流的轨迹预测方法进行对比。结果表明,TPP方法取得了更好的性能表现。

关键词: 车辆轨迹预测, 长时间联合预测, 轨迹后处理, 深度学习, 注意力机制

Abstract:

Precisely perceiving the future actions of surrounding traffic agents is critical for ensuring the safety of autonomous vehicle. This paper mainly focuses on the complicated problem of long-term multi-agent trajectory prediction. Existing trajectory prediction methods can be categorized into joint prediction and marginal prediction. Although joint prediction methods achieve better scene consistency in multi-agent scenarios, both of them fail to reach satisfactory results in long-term prediction tasks due to their inability to capture changes in driver behavior over time. In this paper, we introduce a novel joint prediction method called trajectory prediction through post-processing (TPP). The core of this method is the trajectory post-processing module, which utilizes attention mechanisms to model interactions among different vehicles. By representing a single vehicle with multiple embeddings, the module also simulates the changes in behavior during driving. With the help of a post-processing module, our method is able to generate consistent multi-agent trajectories. We evaluate TPP on a short-term prediction dataset and a long-term prediction dataset separately, comparing it with mainstream trajectory prediction methods. The results indicate that TPP achieves better performance.

Key words: vehicle trajectory prediction, long-term joint prediction, trajectory post-processing, deep learning, attention-mechanism

中图分类号: