Scholarlink

<p>Pedestrian trajectory prediction is essential for autonomous driving and intelligent transportation systems, but existing methods face critical limitations: recurrent models generate trajectories step-by-step, limiting real-time deployment, while current approaches inadequately model multimodal spatiotemporal interactions and lack explicit goal conditioning. To address these challenges, we propose HiGoalNet, a Hierarchical Goal-Conditioned Network designed for real-time trajectory prediction. Our key innovation is a Hierarchical Trajectory Decoder that progressively refines predictions from coarse keypoints to full temporal resolution through multi-scale deconvolutional upsampling in a single forward pass, enabling parallel trajectory generation and eliminating the sequential dependency of recurrent models. Additionally, we develop a Spatiotemporal Multimodal Attention mechanism that captures dynamic interactions across spatial and temporal dimensions, and a Goal-Conditioned Prior Network that incorporates hierarchical planning information. Extensive experiments on ETH and UCY benchmarks demonstrate that HiGoalNet achieves competitive accuracy with average ADE of 0.20 and FDE of 0.34, while requiring only 6.7 ms inference time and 0.6 hours training time, making it particularly suitable for real-time deployment in autonomous systems.</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

HiGoalNet: hierarchical goal-conditioned network with multimodal attention for real-time pedestrian trajectory prediction

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Wenxun Zheng,
  • Wenting Li,
  • Yongjun Zhang

摘要

Pedestrian trajectory prediction is essential for autonomous driving and intelligent transportation systems, but existing methods face critical limitations: recurrent models generate trajectories step-by-step, limiting real-time deployment, while current approaches inadequately model multimodal spatiotemporal interactions and lack explicit goal conditioning. To address these challenges, we propose HiGoalNet, a Hierarchical Goal-Conditioned Network designed for real-time trajectory prediction. Our key innovation is a Hierarchical Trajectory Decoder that progressively refines predictions from coarse keypoints to full temporal resolution through multi-scale deconvolutional upsampling in a single forward pass, enabling parallel trajectory generation and eliminating the sequential dependency of recurrent models. Additionally, we develop a Spatiotemporal Multimodal Attention mechanism that captures dynamic interactions across spatial and temporal dimensions, and a Goal-Conditioned Prior Network that incorporates hierarchical planning information. Extensive experiments on ETH and UCY benchmarks demonstrate that HiGoalNet achieves competitive accuracy with average ADE of 0.20 and FDE of 0.34, while requiring only 6.7 ms inference time and 0.6 hours training time, making it particularly suitable for real-time deployment in autonomous systems.