第50卷第4期2023年4月Vol.50,No.4Apr.2023湖南大学学报(自然科学版)JournalofHunanUniversity(NaturalSciences)基于融合Dropout与注意力机制的LSTM-GRU车辆轨迹预测吴晓建1,2†,危一华1,2,王爱春2,雷耀2,张瑞雪2(1.南昌大学先进制造学院,江西南昌330031;2.江铃汽车股份有限公司,江西南昌,330001)摘要:在智能驾驶环境的车辆轨迹预测环节,为更好地获取环境车辆的轨迹时序特征,在长短期记忆神经网络(LSTM)基础上,嵌入Dropout层以增强网络泛化性,引入注意力机制予以预测效果影响较大的时序数据更大权重从而提高预测结果的可靠性,且将改进的LSTM模型与门控循环单元GRU模型结合,构建LSTM-GRU预测模型以进一步提升环境车辆轨迹预测的准确性.在此基础上,使用NGSIM公开数据集对模型进行训练、验证和测试.研究结果表明,融合了Dropout和注意力机制的LSTM-GRU神经网络轨迹预测模型相较标准的LSTM长短期记忆网络以及GRU门控循环单元,在预测较长时序的车辆轨迹上具有优势,提高了轨迹预测的准确性,降低了实际轨迹和预测轨迹之间的均方根误差和平均绝对误差.关键词:智能汽车;轨迹预测;长短期记忆神经网络;门控循环单元;注意力机制中图分类号:U461.91文献标志码:AVehicleTrajectoryPredictionBasedonLSTM-GRUIntegratingDropoutandAttentionMechanismWUXiaojian1,2†,WEIYihua1,2,WANGAichun2,LEIYao2,ZHANGRuixue2(1.SchoolofAdvancedManufacturing,NanchangUniversity,Nanchang330031,China;2.JianglingMotorsCo.,Ltd.,Nanchang330001,China)Abstract:Intheenvironmentalvehicletrajectorypredictionlinkofintelligentdriving,toobtainthetrajectorytimingcharacteristicsofthesurroundingvehiclesmoreaccurately,thedropoutlayerisembeddedintothelongshort-termmemory(LSTM)neuralnetworksmodeltoenhancenetworkgeneralization,andtheattentionmechanismisintroducedintoLSTMtodistributelargerweighttothetimeseriesdatawithamarkedinfluenceonthepredictioneffect,soastoimprovethereliabilityofthepredictionresults.Subsequently,theimprovedLSTMmodeliscombinedwiththeGRUmodel(LSTM-GRU)tofurtherimprovethepredictionaccuracyofthesurroundingvehicletrajectory.Onthisbasis,theLSTM-GRUmodelistrained,verified,andtestedusingtheNGSIMpublicdataset.Theresultsshowthat,comparedwi...