Scholarlink

<p>Vehicle trajectory prediction is a key aspect of autonomous driving systems, aiming to infer the future movement of vehicles based on historical trajectories and environmental information. Compared to traditional physics- and rule-based methods, deep learning approaches have demonstrated strong performance in modeling complex spatiotemporal interaction patterns from large-scale datasets and can generate Multimodal Trajectory (MT) hypotheses via probabilistic modeling to represent behavioral uncertainty. This paper systematically reviews the research progress of deep learning in trajectory prediction, focusing on the main thread of "task definition, input representation, interaction modeling, map fusion, multimodal generation, evaluation system, engineering application." First, this paper elaborates on task modeling and output forms; next, this paper summarizes methods such as sequence models, interaction modeling, graph networks, Transformers, and generative models; then, this paper reviews mainstream datasets and evaluation metrics, discussing the importance of closed-loop testing, Probabilistic Calibration (PC) and standardized evaluation protocols for reproducibility; finally, we analyze open issues such as out-of-distribution generalization, long-tail risk, and causal interactions, and highlight trends in foundational models, self-supervised learning, and safe verifiable prediction.</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

Deep Learning for Vehicle Trajectory Prediction: A Comprehensive Review of Task Formulation, Spatiotemporal Interaction, Multimodal Generation, and Evaluation Frameworks

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Yi Mei,
  • Bin Liu,
  • Wenxin Li,
  • Yuhonghao Wang

摘要

Vehicle trajectory prediction is a key aspect of autonomous driving systems, aiming to infer the future movement of vehicles based on historical trajectories and environmental information. Compared to traditional physics- and rule-based methods, deep learning approaches have demonstrated strong performance in modeling complex spatiotemporal interaction patterns from large-scale datasets and can generate Multimodal Trajectory (MT) hypotheses via probabilistic modeling to represent behavioral uncertainty. This paper systematically reviews the research progress of deep learning in trajectory prediction, focusing on the main thread of "task definition, input representation, interaction modeling, map fusion, multimodal generation, evaluation system, engineering application." First, this paper elaborates on task modeling and output forms; next, this paper summarizes methods such as sequence models, interaction modeling, graph networks, Transformers, and generative models; then, this paper reviews mainstream datasets and evaluation metrics, discussing the importance of closed-loop testing, Probabilistic Calibration (PC) and standardized evaluation protocols for reproducibility; finally, we analyze open issues such as out-of-distribution generalization, long-tail risk, and causal interactions, and highlight trends in foundational models, self-supervised learning, and safe verifiable prediction.