基于Transformer模型的车辆轨迹预测  

Vehicle Trajectory Prediction Based on Transformer Model

作　　者：田晟[1] 胡啸 TIAN Sheng;HU Xiao(School of Civil Engineering and Transportation,South China University of Technology,Guangzhou Guangdong 510641,China)

机构地区：[1]华南理工大学土木与交通学院,广东广州510641

出　　处：《广西师范大学学报（自然科学版）》2024年第3期47-58,共12页Journal of Guangxi Normal University:Natural Science Edition

基　　金：广东省自然科学基金(2021A1515011587)。

摘　　要：准确预测车辆轨迹可以保障自动驾驶车辆行驶安全,针对已有方法对长序列轨迹建模预测能力有限的问题,本文提出一种基于Transformer网络的车辆轨迹预测模型。将车辆的运动数据与交互数据输入驾驶意图预测模块生成概率意图向量,通过Concatenate函数与轨迹信息拼接后输入轨迹预测编码器,利用多头注意力机制充分提取轨迹特征,经解码器得到未来时刻的车辆轨迹分布。在车辆轨迹真实数据集NGSIM上进行验证,结果表明:在2 s预判时间下,驾驶意图预测模块准确率可达到85%以上;在4 s的预测时域下,轨迹预测模型相较于已有模型,其RMSE降低均达到10%以上。本文提出方法为自动驾驶车辆准确预测轨迹提供技术支持。Accurately predicting the trajectory of vehicle is crucial to ensure the safety of autonomous vehicles.However,traditional methods have limited modeling and predictive capabilities when dealing with long sequence trajectories.To address this issue,a vehicle trajectory prediction model was proposed based on the Transformer network.The approach involves inputting the motion and interaction data of the vehicle into a driving intention prediction module to generate a probability intention vector.The trajectory prediction encoder is obtained after the Concatenate function is spliced with the trajectory information,and the trajectory features are fully extracted by using the multi-head attention mechanism.Through the decoder,a distribution of future vehicle trajectories is obtained.Validation on the NGSIM real vehicle trajectory dataset indicates that the accuracy of the driving intention prediction module can reach more than 85%under a prediction time of 2 seconds.Furthermore,the RMSE of the trajectory prediction model is reduced by more than 10%compared with the existing models under a prediction time of 4 seconds.The method provides technical support for accurately predicting the trajectory of autonomous vehicles.

关 键 词：自动驾驶 车辆轨迹预测 驾驶意图 特征提取 多头注意力机制 

分 类 号：U495[交通运输工程—交通运输规划与管理]