基于边更新与多头交互融合Transformer的车辆轨迹预测方法

打开文本图片集

关键词：车辆轨迹预测；深度学习；Transformer；多头注意力机制；时空特征融合；智能驾驶中图分类号：TP13 文献标志码：A 文章编号：1001-3695（2025）08-014-2348-07doi：10.19734/j. issn. 1001-3695.2025.01.0017

Vehicles trajectory prediction approach based on Transformer with edge update and multi-head attention interactive fusion

Sun Yinga†，Wu Yanyongʰ，Ding Deruib，Zhang Jiankunb （a.BusinessSolholfpticalElecrical&ompuerEgineig，Uiesitfgifoice&echolog ，China）

Abstract：Thetaskofvehicletrajectorypredictionforautonomous driving needs tofullyconsider therelationship betweenthe traficagentsandtheenvironment.Addresingthelimitationsofexistingapproachesatthelevelofheterogeneousfeaturenteractionand improving prediction accuracy，the paper proposedavehicle trajectoryprediction approach named EMATNet with edgeupdatingandmulti-headattention interactivefusionTransformer.Firstly，theapproach encodedand embeddedthe historicalspatio-temporal informationof theagentsandthetransportationenvironment.Then，theapproachusedtheproposedtwostageinteractionnetworkofedgeupdatingandmulti-atentioninteraction fusion Transformerforfeature interaction.The introducedsymmetricpositionalembeddingandvehicle-roadrelationshipinteractioncouldefectivelyenhancetheglobalinformation perceptionandspatio-temporalrelationshipcapturingcapability.Finalythisapproachusedtwostageoptimizationdecoding to ensurethe acuracyand reasonablenessofthe predictionresults.The proposedappoachvalidatedonArgoverseland Argoverse2 motion prediction datasets，and visualizedandanalyzed the predictionresults.Theresults show that EMATNetoutperforms similarapproaches inthethre performance metricsofminFDE，minADEandMR，andiscapableforthe taskof vehicle trajectory prediction in complex traffic environments.

Keywords：vehicletrajectoryprediction；deeplearning;Transformer;multi-headatentionmechanism；spatio-temporalfeature integration； intelligent driving

0 引言

近年来，智能驾驶技术正吸引着学术界和工业界越来越多的关注，其在交通领域中有着越来越广泛的应用。（剩余17833字）