Scholarlink

<p>Target vehicle trajectory prediction is of great significance in the field of autonomous driving. To achieve accurate trajectory prediction in an autonomous driving system, it is necessary to effectively model the complex spatio-temporal interactions between dynamic entities (vehicles, pedestrians) and static infrastructure. With the increasing complexity of urban traffic scenarios, existing methods face three core challenges: (1) The dynamically evolving characteristics of spatial dependencies affected by random factors such as traffic accidents and driving behaviors; (2) Multi-scale time patterns covering instantaneous maneuvers to long-term trends; (3) Severe constraints on real-time computational efficiency in high-density traffic scenarios. Therefore, we propose a real-time vehicle trajectory prediction method based on the fusion of spatio-temporal awareness graph and multi-scale dilated convolution. This method dynamically constructs the vehicle-lane interaction topology to solve the problem of high spatial modeling errors of traditional static graph methods in scenarios such as sudden lane changes; captures microscopic trajectory fluctuations and macroscopic motion trends simultaneously through multi-level dilation rate spatio-temporal dilated convolution; introduces Top-k screening focused additive attention to achieve the interaction modeling accuracy of the full attention mechanism with a significant inference speed in vehicle scenarios. Experimental results show that our method is significantly superior to existing methods on the Argoverse dataset, demonstrating its great potential in practical applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A real-time vehicle trajectory prediction method based on the fusion of spatio-temporal awareness graph and multi-scale dilated convolution

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Xiang Gu,
  • Chenwen Gu,
  • Jing Wang,
  • Chao Li,
  • Qiwei Huang

摘要

Target vehicle trajectory prediction is of great significance in the field of autonomous driving. To achieve accurate trajectory prediction in an autonomous driving system, it is necessary to effectively model the complex spatio-temporal interactions between dynamic entities (vehicles, pedestrians) and static infrastructure. With the increasing complexity of urban traffic scenarios, existing methods face three core challenges: (1) The dynamically evolving characteristics of spatial dependencies affected by random factors such as traffic accidents and driving behaviors; (2) Multi-scale time patterns covering instantaneous maneuvers to long-term trends; (3) Severe constraints on real-time computational efficiency in high-density traffic scenarios. Therefore, we propose a real-time vehicle trajectory prediction method based on the fusion of spatio-temporal awareness graph and multi-scale dilated convolution. This method dynamically constructs the vehicle-lane interaction topology to solve the problem of high spatial modeling errors of traditional static graph methods in scenarios such as sudden lane changes; captures microscopic trajectory fluctuations and macroscopic motion trends simultaneously through multi-level dilation rate spatio-temporal dilated convolution; introduces Top-k screening focused additive attention to achieve the interaction modeling accuracy of the full attention mechanism with a significant inference speed in vehicle scenarios. Experimental results show that our method is significantly superior to existing methods on the Argoverse dataset, demonstrating its great potential in practical applications.