Scholarlink

<p>Vehicle trajectory prediction is a critical component of autonomous driving technology, aiming to predict plausible future trajectories for surrounding agents in dynamic traffic scenarios. However, recent approaches merely fit certain elements of interest using the agent’s local poses, which is insufficient for effectively perceiving large-scale static maps. Meanwhile, some models rely on historical or current map states, assigning excessive attention to traversed road segments. This leads to severe limitations in vehicles’ autonomous decision-making capabilities. Furthermore, single-stage prediction fails to evaluate intersecting future spatial distributions, consequently inducing hazardous behaviors. To address these challenges, this paper proposes a Future Occupancy Guidance(FOG) framework based on multi-view collaboration. Firstly, the model leverages edge relationships to construct spatio-temporal interactions among elements across multiple viewpoint graphs. Then, we generate future occupancy markers along the driving direction and adaptively shift the map query perspective to accommodate the distance disparities caused by agents traveling at different speeds, thereby providing road-level guidance. Finally, we design a refinement module to further correct the rough predicted trajectories with collision risks through multi-head attention. The experimental results on Argoverse 1 and Argoverse 2 motion forecasting benchmarks demonstrate that our FOG achieves outstanding performance in reducing both FDE and MR metrics. The code implementation is available at: <a href="https://github.com/alien-HL/FOG.git">https://github.com/alien-HL/FOG.git</a></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

Future occupancy guidance under multi-view collaboration for trajectory prediction

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Haifeng Sang,
  • Lai Huang

摘要

Vehicle trajectory prediction is a critical component of autonomous driving technology, aiming to predict plausible future trajectories for surrounding agents in dynamic traffic scenarios. However, recent approaches merely fit certain elements of interest using the agent’s local poses, which is insufficient for effectively perceiving large-scale static maps. Meanwhile, some models rely on historical or current map states, assigning excessive attention to traversed road segments. This leads to severe limitations in vehicles’ autonomous decision-making capabilities. Furthermore, single-stage prediction fails to evaluate intersecting future spatial distributions, consequently inducing hazardous behaviors. To address these challenges, this paper proposes a Future Occupancy Guidance(FOG) framework based on multi-view collaboration. Firstly, the model leverages edge relationships to construct spatio-temporal interactions among elements across multiple viewpoint graphs. Then, we generate future occupancy markers along the driving direction and adaptively shift the map query perspective to accommodate the distance disparities caused by agents traveling at different speeds, thereby providing road-level guidance. Finally, we design a refinement module to further correct the rough predicted trajectories with collision risks through multi-head attention. The experimental results on Argoverse 1 and Argoverse 2 motion forecasting benchmarks demonstrate that our FOG achieves outstanding performance in reducing both FDE and MR metrics. The code implementation is available at: https://github.com/alien-HL/FOG.git