Scholarlink

<p>Car-following is a fundamental driving behavior that plays a crucial role in maintaining traffic safety. To enhance the accuracy and robustness of car-following prediction in autonomous driving, we propose a novel model: the Bidirectional Transformer attention mechanism for Car following (BiTAM-CF). The model employs a Bidirectional Long Short-Term Memory (BiLSTM) network as a temporal encoder to capture long-term dependencies from historical states, and leverages the Transformer’s global modeling capabilities to capture interactions across time steps and feature dimensions. To address the dynamic characteristics of real-world traffic scenarios, we introduce two mechanisms: Adaptive Multi-Scale Attention (AMSA) and Spatiotemporal Attention (STA). AMSA captures behavioral variations across multiple temporal scales, whereas STA models critical driving intentions and dynamic behavioral changes. Experiments conducted on five public datasets, including NGSIM, HighD, and Waymo, demonstrate that BiTAM-CF consistently outperforms baseline models in prediction accuracy. For instance, on the NGSIM dataset, BiTAM-CF reduces speed prediction error by approximately 35% compared to other models, while on the Waymo dataset, spacing error is reduced by about 15%. These results suggest that the proposed BiTAM-CF achieves accurate and robust car-following predictions across diverse scenarios.</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 Spatiotemporal BiLSTM-Transformer Model with Multi-Scale Attention for Car-Following Prediction

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • Weiwei Huo,
  • Jintao Liu,
  • Zehui Zhang

摘要

Car-following is a fundamental driving behavior that plays a crucial role in maintaining traffic safety. To enhance the accuracy and robustness of car-following prediction in autonomous driving, we propose a novel model: the Bidirectional Transformer attention mechanism for Car following (BiTAM-CF). The model employs a Bidirectional Long Short-Term Memory (BiLSTM) network as a temporal encoder to capture long-term dependencies from historical states, and leverages the Transformer’s global modeling capabilities to capture interactions across time steps and feature dimensions. To address the dynamic characteristics of real-world traffic scenarios, we introduce two mechanisms: Adaptive Multi-Scale Attention (AMSA) and Spatiotemporal Attention (STA). AMSA captures behavioral variations across multiple temporal scales, whereas STA models critical driving intentions and dynamic behavioral changes. Experiments conducted on five public datasets, including NGSIM, HighD, and Waymo, demonstrate that BiTAM-CF consistently outperforms baseline models in prediction accuracy. For instance, on the NGSIM dataset, BiTAM-CF reduces speed prediction error by approximately 35% compared to other models, while on the Waymo dataset, spacing error is reduced by about 15%. These results suggest that the proposed BiTAM-CF achieves accurate and robust car-following predictions across diverse scenarios.