Scholarlink

<p>Accurately predicting the future trajectories of surrounding vehicles is essential for autonomous vehicles to make proactive decisions in dynamic, human-dominated traffic environments. This task is challenging due to the complex interactions between vehicles, the unpredictable human driving behavior, and the need to account for both temporal patterns and social interactions. Inspired by the recent advancements in deep learning, particularly with transformers and LSTMs, in this paper, we propose a new spatial-temporal transformer architecture for maneuver-aware trajectory prediction. Our model employs dedicated spatial and temporal transformer modules to capture dynamic dependencies between the target vehicle and its neighbors. We design a spatial embedding framework to ensure dimensional consistency by masking positions without neighboring vehicles. We further design a decoder that includes a temporal transformer module that can integrate embeddings generated by the encoder in addition to the target vehicle’s original trajectory. Our decoder enables accurate long-term predictions while maintaining short-term precision. Experimental results on the NGSIM US-101 and I-80 datasets demonstrate that this approach achieves improved performance in long-term trajectory prediction and can enhance the safety and efficiency of autonomous driving 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

Maneuver-based Trajectory Prediction for Automated Driving on Highways Using Spatial and Temporal Transformers

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Asmaa Loulou,
  • Mustafa Unel

摘要

Accurately predicting the future trajectories of surrounding vehicles is essential for autonomous vehicles to make proactive decisions in dynamic, human-dominated traffic environments. This task is challenging due to the complex interactions between vehicles, the unpredictable human driving behavior, and the need to account for both temporal patterns and social interactions. Inspired by the recent advancements in deep learning, particularly with transformers and LSTMs, in this paper, we propose a new spatial-temporal transformer architecture for maneuver-aware trajectory prediction. Our model employs dedicated spatial and temporal transformer modules to capture dynamic dependencies between the target vehicle and its neighbors. We design a spatial embedding framework to ensure dimensional consistency by masking positions without neighboring vehicles. We further design a decoder that includes a temporal transformer module that can integrate embeddings generated by the encoder in addition to the target vehicle’s original trajectory. Our decoder enables accurate long-term predictions while maintaining short-term precision. Experimental results on the NGSIM US-101 and I-80 datasets demonstrate that this approach achieves improved performance in long-term trajectory prediction and can enhance the safety and efficiency of autonomous driving systems.