Scholarlink

<p>Pedestrian trajectory prediction aims to use observed human historical trajectories and surrounding environmental information to forecast the future positions of pedestrians. This research is particularly valuable in fields like autonomous driving. Existing methods typically use generative approaches, which are repeatedly sampling multiple latent variables in a potential space to represent the diversity of trajectories. Although these methods are effective, the latent variables pose interpretability issues. Furthermore, most approaches predict multimodal trajectories without incorporating probability information, which is crucial for safe decision-making in autonomous driving. In this paper, we introduce a framework PTMP (predefined trajectories for multimodal pedestrian trajectory prediction), which screens through a large volume of historical trajectories to generate ‘Pedestrian Walking Modes’ that allow for interpretable pedestrian trajectory predictions. PTMP employs a dual-prediction strategy to forecast diverse trajectories and their corresponding probabilities, addressing the multimodality of future trajectories. Extensive experiments on ETH and UCY datasets demonstrate that PTMP outperforms state-of-the-art methods, achieving an average ADE of 0.33 and FDE of 0.56, improving upon STGAT by 23.3% and 32.5%, respectively. Ablation studies confirm the effectiveness of each PTMP component. Our code is publicly available at <a href="https://github.com/tengZ222/PTMP">https://github.com/tengZ222/PTMP</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

PTMP: predefined trajectories for multimodal pedestrian trajectory prediction

  • 期刊论文

  • 发表日期：

    2025年6月20日

查看全文
  • Teng Zhang,
  • Bo Yang,
  • Jianlin Zhu,
  • Xincheng Hu

摘要

Pedestrian trajectory prediction aims to use observed human historical trajectories and surrounding environmental information to forecast the future positions of pedestrians. This research is particularly valuable in fields like autonomous driving. Existing methods typically use generative approaches, which are repeatedly sampling multiple latent variables in a potential space to represent the diversity of trajectories. Although these methods are effective, the latent variables pose interpretability issues. Furthermore, most approaches predict multimodal trajectories without incorporating probability information, which is crucial for safe decision-making in autonomous driving. In this paper, we introduce a framework PTMP (predefined trajectories for multimodal pedestrian trajectory prediction), which screens through a large volume of historical trajectories to generate ‘Pedestrian Walking Modes’ that allow for interpretable pedestrian trajectory predictions. PTMP employs a dual-prediction strategy to forecast diverse trajectories and their corresponding probabilities, addressing the multimodality of future trajectories. Extensive experiments on ETH and UCY datasets demonstrate that PTMP outperforms state-of-the-art methods, achieving an average ADE of 0.33 and FDE of 0.56, improving upon STGAT by 23.3% and 32.5%, respectively. Ablation studies confirm the effectiveness of each PTMP component. Our code is publicly available at https://github.com/tengZ222/PTMP.