基于注意力Seq2Seq网络的高速公路交织区车辆变道轨迹预测  

Lane Change Trajectory Prediction of Vehicles in Highway Interweaving Area Using Seq2Seq-attention Network

作　　者：韩皓[1] 谢天 HAN Hao;XIE Tian(College of Transport&Communications,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学交通运输学院,上海201306

出　　处：《中国公路学报》2020年第6期106-118,共13页China Journal of Highway and Transport

基　　金：国家自然科学基金项目(71871137);上海市人民政府决策咨询研究项目(2015ZD16B)。

摘　　要：针对交通状态复杂的高速公路交织区域,经验丰富的驾驶人能够通过正确地推断周围车辆的未来运动进行及时的车道变换,这对于实现安全高效的自动驾驶至关重要,然而目前的自动驾驶车辆往往缺乏这种预测能力。为此,基于深度学习理论,提出了一种结合注意力机制和编-解码器结构的交织区车辆强制性变道轨迹预测方法,利用Next Generation Simulation(NGSIM)数据集提取车辆变道过程中的关键特征,并引入碰撞时间(Time to Collision,TTC)和避免碰撞减速度(Deceleration Rate to Avoid a Crash,DRAC)2种风险指标,将变道车辆及其周围车辆视为一个整体状态单元,同时补全状态单元内部不同车辆在横向和纵向上的时空状态特征,从而更有效地刻画车辆间的动态交互行为;然后将不同观测车辆的连续窗口序列输入基于长短期记忆网络(Long Short-term Memory,LSTM)的编-解码器,预测交织区车辆变道的未来运动轨迹,通过添加软注意力模块,使模型能够集中聚焦于影响车辆在不同时刻下位置变化的关键信息,再现了真实交通场景下车辆的变道行为。试验验证表明:基于注意力机制的编-解码器模型与当前流行的卷积长短期记忆网络、极限梯度提升树等模型相比具有更高的轨迹预测精度,在长时域的变道轨迹拟合上有显著的优越性,为辅助和自动驾驶领域的发展提供了新思路。In highway interweaving areas with complex traffic states,experienced drivers can make timely lane changes by correctly inferring the future movements of surrounding vehicles.This is essential for safe and efficient autonomous driving.However,existing autonomous vehicles often lack this prediction ability.Therefore,in this study,a method is proposed to predict the mandatory lane change trajectories of vehicles in the highway interweaving area.This method combines an Attention Mechanism and sequence-to-sequence(Seq2 Seq)network structure,using next generation simulation(NGSIM)data set to extract key features during vehicle lane-changing process,and introduces two types of risk indicators:time to collision(TTC)and deceleration rate to avoid a crash(DRAC).It treats the lane change vehicle and its surrounding vehicles as an overall state unit and simultaneously completes the spatio-temporal features of different vehicles in the state unit at lateral and longitudinal directions,to effectively describe the dynamic interaction between vehicles.Subsequently,the continuous window sequences of different observed vehicles are input into the Seq2 Seq model,based on the long short-term memory(LSTM)network,to predict the future motion trajectories of lane change vehicles in the interweaving area,by adding a Soft Attention module.This module focuses on key information that affects the position change of vehicles at different times and reproduces the lanechanging behavior of vehicles in real traffic scenarios.The experimental verification results illustrate that the Seq2 Seq-attention model has higher trajectory prediction accuracy than the currently popular models,such as ConvLSTM,XGBoost,especially for long-term horizon trajectory fitting.In addition,it provides new ideas for the development of assisted and autonomous driving.

关 键 词：交通工程 交织区变道 轨迹预测 注意力机制 编-解码器结构 车辆交互 长短期记忆网络 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]