基于多尺度时空注意力机制的车辆换道意图预测

Vehicle Lane Change Intention Prediction Based on Multi-Scale Spatial-Temporal Attention Mechanism

摘要: 车辆换道行为预测是智能交通系统和自动驾驶技术的核心组成部分，对提升道路交通安全具有重要意义。针对现有方法在时空特征建模分离、多尺度信息融合不足和动态权重调整缺失等方面的技术挑战，本文提出了一种基于多尺度时空注意力机制的车辆换道意图预测模型(MSTAN)。该模型采用分治协同的设计思想，构建了三个并行处理分支：时序特征建模分支基于双向长短时记忆网络和自注意力机制捕获车辆运动状态的时间依赖关系；空间交互建模分支利用图卷积网络和边注意力机制建模车辆间的动态空间关系；多尺度特征融合分支通过并行多分辨率卷积和注意力机制提取多粒度上下文信息。模型采用自适应权重融合策略实现三个分支特征的最优整合。基于NGSIM数据集的实验结果表明，MSTAN在US-101、I-80和Lankershim三个路段的换道行为三分类任务中分别达到98.3%、97.6%和95.2%的预测准确率，优于所有基线方法。并通过分支贡献度分析和注意力机制有效性评估，定量验证了各功能模块对模型性能的独立贡献和协同效应。时间窗口敏感性分析和预测时间提前量分析确定了最优的时序建模参数配置，证明了模型在不同预测场景下的稳定性和鲁棒性。

Abstract: Vehicle lane change behavior prediction is a core component of intelligent transportation systems and autonomous driving technologies, playing a crucial role in enhancing road traffic safety. Addressing the technical challenges of existing methods in spatial-temporal feature modeling separation, insufficient multi-scale information fusion, and the absence of dynamic weight adjustment, this paper proposes a Multi-Scale Spatial-Temporal Attention Network (MSTAN) for vehicle lane change intention prediction. The model adopts a divide-and-conquer collaborative design philosophy, constructing three parallel processing branches: the temporal feature modeling branch captures temporal dependencies of vehicle motion states based on bidirectional long short-term memory networks and self-attention mechanisms; the spatial interaction modeling branch utilizes graph convolutional networks and edge attention mechanisms to model dynamic spatial relationships between vehicles; the multi-scale feature fusion branch extracts multi-granularity contextual information through parallel multi-resolution convolutions and attention mechanisms. The model employs an adaptive weight fusion strategy to achieve optimal integration of features from the three branches. Experimental results on the NGSIM dataset demonstrate that MSTAN achieves prediction accuracies of 98.3%, 97.6%, and 95.2% for lane change behavior three-class classification tasks on the US-101, I-80, and Lankershim road segments, respectively, outperforming all baseline methods. Through branch contribution analysis and attention mechanism effectiveness evaluation, the independent contributions and synergistic effects of each functional module on model performance are quantitatively validated. Time window sensitivity analysis and prediction lead time analysis determine the optimal temporal modeling parameter configuration, demonstrating the model’s stability and robustness across different prediction scenarios.