基于軌跡預(yù)測(cè)模型的仿真車(chē)輛軌跡生成算法

打開(kāi)文本圖片集

Trajectory generation algorithm for simulated vehicles based on trajectory prediction models

WANG Zhenyu，YU Zhuoping，TIANWei*，XIONG Lu，LI Zhuoren （School ofAutomotivetudies，TongjiUniversity，hanghail8oo，China）

Abstract：Toenhance theoverallrealismof background interactive vehicle trajectories indigital simulation scenarios forautonomous driving，this studyapproached the problem from both microscopicand macroscopic perspectives.Firstly，vehicle trajectory prediction models were trained onnaturalistic driving data.Leveraging thecharacteristic that model-predicted trajectories more closely resembled real-world vehicle trajectories， the prediction servedas the artificial intelligence （Al） driver model for background vehicles in simulation environments，improving themicroscopic realismof simulatedvehicle trajectory interactions.Building on this foundation，ameasurement method fortrajectory feature parameter statistical distribution differencesand a correspondingoptimizationalgorithmweredesigned，tore-selectasingletrajectorywiththehighestprobability frommultiple multi-modalpredictionoutputs，as the final driving trajectoryforsimulated vehicles，further enhancing the macroscopic realismof the generated trajectory feature parameter statistical distribution.The resultsshow that，based on theproposed measurement metrics，the distributiondiffrence betweenoptimized simulated trajectories and real trajectories is reduced by 56.29% compared to pre-optimization，effectively enhancing the realismof background vehicle trajectories insimulationscenarios.

Keywords：multimodaltrajectoryprediction； trajectorysnapshot；trajectory feature vectorclustering；KullbackLeibler （KL） divergence； Bayesian optimization

自動(dòng)駕駛技術(shù)雖然經(jīng)過(guò)長(zhǎng)時(shí)間發(fā)展，但L3以上高級(jí)別自動(dòng)駕駛汽車(chē)依然面臨無(wú)法商業(yè)化落地的難題。（剩余11668字）