混合交通流环境下基于时变交互关系的车辆轨迹预测方法研究

Research on Vehicle Trajectory Prediction Method Based on Time-Varying Interaction Relationship in Mixed Traffic Flow Environment

摘要: 在自动驾驶车辆(AV)与有人驾驶车辆(HDV)长期混行的现实道路环境中，车辆之间的交互关系具有高度的时变性和语义依赖性。传统基于静态邻接矩阵或固定阈值的交互建模方法在面对变道、加塞、紧急制动等突发场景时，往往无法快速反映交互关系的变化，导致预测精度和鲁棒性下降。为此，本文提出一种基于时变交互关系建模方法，核心包括三部分：(1) 利用物理特征(空间距离、相对速度、航向差、加速度差)构建多因素耦合的基础邻接权重函数；(2) 根据指数平滑与门控机制对邻接矩阵进行时序更新，从而获得时变邻接矩阵；(3) 将时变邻接矩阵与车辆历史状态一同输入轻量级GRU时序网络进行未来多模态轨迹预测。在NuScenes公开数据集上的对比实验表明，所提方法在平均位移ADE、MAE等指标上均优于静态图方法，并在多种典型场景(并线、跟车、拥堵起步)中表现出更好的响应性和稳定性。

Abstract: In the real road environment where autonomous vehicles (AVs) and human-driven vehicles (HDVs) coexist for a long time, the interaction relationships among vehicles are highly time-varying and semantically dependent. Traditional interaction modeling methods based on static adjacency matrices or fixed thresholds often fail to rapidly reflect changes in interaction relationships when facing sudden scenarios such as lane changing, cutting in, and emergency braking, leading to degraded prediction accuracy and robustness. To this end, this paper proposes a modeling method based on time-varying interaction relationships, which mainly consists of three parts: (1) constructing a multi-factor coupled basic adjacency weight function using physical features including spatial distance, relative velocity, heading difference, and acceleration difference; (2) updating the adjacency matrix sequentially via exponential smoothing and gating mechanism to obtain a time-varying adjacency matrix; (3) feeding the time-varying adjacency matrix together with vehicle historical states into a lightweight GRU temporal network for future multimodal trajectory prediction. Comparative experiments on the public NuScenes dataset show that the proposed method outperforms static graph-based methods in terms of average displacement error (ADE), mean absolute error (MAE) and other metrics, and exhibits better responsiveness and stability in various typical scenarios such as lane merging, car following, and congestion start.