考虑环境与驾驶习惯的车辆加速预测,提升智能驾驶控制精度
Vehicle Acceleration Prediction Considering Environmental Influence and Individual Driving Behavior
- 双输入设计:融合环境交通数据与个体驾驶行为序列
- 引入司机分类后,预测准确率提升33%
- 适用于需要个性化驾驶策略的智能交通系统
精准的车辆加速预测对智能驾驶控制与能效管理至关重要,尤其在复杂动态驾驶环境中。本文提出一种通用的短期车辆加速预测框架,联合建模环境影响与个体驾驶行为。框架采用双输入设计:环境序列由特定空间位置的历史交通变量(如分位数速度、多车加速度统计)构建,捕捉群体驾驶行为受交通环境的影响;个体驾驶行为序列则反映目标车辆在预测点前的运动特征,体现个性化驾驶风格。两个输入通过增强注意力机制的LSTM Seq2Seq模型处理,实现多步加速预测。基于广州白石隧道出口段高分辨率雷达视频融合轨迹数据进行实证研究,依据关键行为指标将驾驶员分为保守、中等和激进三类,并为每类训练专属预测模型以应对司机异质性。实验结果表明,所提方法持续优于四种基线模型,在引入历史交通变量时准确率提升10.9%,结合司机分类后提升达33%。尽管预测误差随预报距离增加,但融入环境与行为感知特征显著增强了模型鲁棒性。
原文摘要 · Abstract (English)
Accurate vehicle acceleration prediction is critical for intelligent driving control and energy efficiency management, particularly in environments with complex driving behavior dynamics. This paper proposes a general short-term vehicle acceleration prediction framework that jointly models environmental influence and individual driving behavior. The framework adopts a dual input design by incorporating environmental sequences, constructed from historical traffic variables such as percentile-based speed and acceleration statistics of multiple vehicles at specific spatial locations, capture group-level driving behavior influenced by the traffic environment. In parallel, individual driving behavior sequences represent motion characteristics of the target vehicle prior to the prediction point, reflecting personalized driving styles. These two inputs are processed using an LSTM Seq2Seq model enhanced with an attention mechanism, enabling accurate multi-step acceleration prediction. To demonstrate the effectiveness of the proposed method, an empirical study was conducted using high resolution radar video fused trajectory data collected from the exit section of the Guangzhou Baishi Tunnel. Drivers were clustered into three categories conservative, moderate, and aggressive based on key behavioral indicators, and a dedicated prediction model was trained for each group to account for driver heterogeneity.Experimental results show that the proposed method consistently outperforms four baseline models, yielding a 10.9% improvement in accuracy with the inclusion of historical traffic variables and a 33% improvement with driver classification. Although prediction errors increase with forecast distance, incorporating environment- and behavior-aware features significantly enhances model robustness.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品;中文卡片由大模型生成,请以原文为准。