Scholarlink

<p>Accurate prediction of vehicle trajectories and driving intentions is critical for ensuring safety and efficiency in complex human–machine hybrid traffic environments. However, existing graph-based methods often rely on heuristic or fully learned interaction structures, which may introduce redundant connections and lack causal interpretability. To address this limitation, this paper proposes a novel framework, termed the Causality-Augmented Graph Convolutional Network (CA-GCN), which integrates causal inference with graph neural networks for lane-change intention prediction. Specifically, Granger causality analysis is employed to uncover predictive dependencies among trajectory variables, from which a directed sparse causal graph is constructed as a structural prior. This prior constrains the topology of the graph convolutional network, while allowing edge weights to be adaptively optimized during training to capture residual non-linear interactions. Furthermore, to ensure methodological consistency, we analyze the relationship between mutual information-based feature selection and Granger causality filtering, showing that the majority of informative features are preserved after causal screening. Based on the learned spatiotemporal representations, the proposed framework performs joint trajectory prediction and driving intention classification, including lane keeping, left lane change, and right lane change. Extensive experiments on the NGSIM and HighD datasets demonstrate that, under a 5-second observation and 3-second prediction horizon, CA-GCN achieves a trajectory prediction RMSE of 1.3 m, significantly outperforming the baseline GCN (1.85 m). For intention recognition, the model attains accuracy and sensitivity exceeding 96%, indicating strong robustness under imbalanced traffic scenarios. The results validate that incorporating a causality-driven structural prior into graph neural networks improves both prediction accuracy and interpretability, providing a reliable and physically grounded solution for vehicle behavior understanding in complex traffic environments.</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 causal inference-integrated graph neural network model for lane-changing intent prediction of vehicle driving

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月15日

查看全文
  • Gaoteng Yuan,
  • Ping Qiu,
  • Jun Zhuang,
  • Xiang Li,
  • Wanli Feng

摘要

Accurate prediction of vehicle trajectories and driving intentions is critical for ensuring safety and efficiency in complex human–machine hybrid traffic environments. However, existing graph-based methods often rely on heuristic or fully learned interaction structures, which may introduce redundant connections and lack causal interpretability. To address this limitation, this paper proposes a novel framework, termed the Causality-Augmented Graph Convolutional Network (CA-GCN), which integrates causal inference with graph neural networks for lane-change intention prediction. Specifically, Granger causality analysis is employed to uncover predictive dependencies among trajectory variables, from which a directed sparse causal graph is constructed as a structural prior. This prior constrains the topology of the graph convolutional network, while allowing edge weights to be adaptively optimized during training to capture residual non-linear interactions. Furthermore, to ensure methodological consistency, we analyze the relationship between mutual information-based feature selection and Granger causality filtering, showing that the majority of informative features are preserved after causal screening. Based on the learned spatiotemporal representations, the proposed framework performs joint trajectory prediction and driving intention classification, including lane keeping, left lane change, and right lane change. Extensive experiments on the NGSIM and HighD datasets demonstrate that, under a 5-second observation and 3-second prediction horizon, CA-GCN achieves a trajectory prediction RMSE of 1.3 m, significantly outperforming the baseline GCN (1.85 m). For intention recognition, the model attains accuracy and sensitivity exceeding 96%, indicating strong robustness under imbalanced traffic scenarios. The results validate that incorporating a causality-driven structural prior into graph neural networks improves both prediction accuracy and interpretability, providing a reliable and physically grounded solution for vehicle behavior understanding in complex traffic environments.