Scholarlink

<p>Accurate trajectory prediction is essential for autonomous vehicles to navigate safely without causing harm to other traffic participants. Many existing deep learning models rely on homogeneous graphs, which fail to capture the complex interactions among different types of road users, such as vehicles, pedestrians, and cyclists. This paper introduces ROTA, a novel trajectory prediction model that leverages heterogeneous dynamic relation-awareness graphs combined with graph neural networks and a transformer-based feature importance assignment mechanism. By integrating spatial–temporal relationships dynamically, ROTA provides a more comprehensive modeling of diverse object interactions. The effectiveness of ROTA is evaluated using the Waymo Open Perception dataset ground-truth data, demonstrating a 27% and 32% reduction in average displacement error and final displacement error, respectively, compared to state-of-the-art methods. Additionally, this study explores the impact of different time series processing techniques, including Conv1D, long short-term memory, and gated recurrent unit, highlighting the superior performance of recurrent networks in capturing motion patterns. Furthermore, an analysis of dataset biases and their influence on model generalizability is provided. These findings suggest that ROTA presents a promising approach for enhancing the predictive capabilities 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

Spatial–Temporal Predictor of Future Trajectories of Traffic Objects with Dynamic Relation-Awareness Heterogeneous Graph for Autonomous Cars: ROTA

  • 期刊论文

  • 发表日期：

    2025年9月29日

查看全文
  • Sevcan Turan,
  • Feyzullah Temurtaş,
  • Mehmet Tektaş

摘要

Accurate trajectory prediction is essential for autonomous vehicles to navigate safely without causing harm to other traffic participants. Many existing deep learning models rely on homogeneous graphs, which fail to capture the complex interactions among different types of road users, such as vehicles, pedestrians, and cyclists. This paper introduces ROTA, a novel trajectory prediction model that leverages heterogeneous dynamic relation-awareness graphs combined with graph neural networks and a transformer-based feature importance assignment mechanism. By integrating spatial–temporal relationships dynamically, ROTA provides a more comprehensive modeling of diverse object interactions. The effectiveness of ROTA is evaluated using the Waymo Open Perception dataset ground-truth data, demonstrating a 27% and 32% reduction in average displacement error and final displacement error, respectively, compared to state-of-the-art methods. Additionally, this study explores the impact of different time series processing techniques, including Conv1D, long short-term memory, and gated recurrent unit, highlighting the superior performance of recurrent networks in capturing motion patterns. Furthermore, an analysis of dataset biases and their influence on model generalizability is provided. These findings suggest that ROTA presents a promising approach for enhancing the predictive capabilities of autonomous driving systems.