智能网联车辆MFP算法轨迹预测模型研究  

Research on MFP algorithm-based trajectory prediction model for intelligent connected vehicles

作　　者：何博[1,2,3] 黄妙华[1,2,3] 刘若璎 邹天越 尹思源 胡永康 HE Bo;HUANG Miaohua;LIU Ruoying;ZOU Tianyue;YIN Siyuan;HU Yongkang(Hubei Key Laboratory of Advanced Technology for Automotive Components,Wuhan University of Technology,Wuhan 430070,China;Hubei Collaborative Innovation Center for Automotive Components Technology,Wuhan University of Technology,Wuhan 430070,China;Hubei Research Center for New Energy&Intelligent Connected Vehicle,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学现代汽车零部件技术湖北省重点实验室,武汉430070 [2]武汉理工大学汽车零部件技术湖北省协同创新中心,武汉430070 [3]武汉理工大学湖北省新能源与智能网联车工程技术研究中心,武汉430070

出　　处：《重庆理工大学学报（自然科学）》2024年第10期21-27,共7页Journal of Chongqing University of Technology：Natural Science

基　　金：国家重点研发计划(2018YFE0105500)。

摘　　要：为解决车辆轨迹预测任务中环境不断变化、车辆间存在交互影响,导致长期预测情况下预测精度较低的问题,提出了一种基于multiple futures predictor(MFP)算法的多智能体轨迹预测模型。采用对称指数移动平均法去除异常数据并平滑轨迹;采用图神经网络(graph convolutional neural network,GCN)进行交互特征提取,将历史轨迹与未来智能体之间的交互特征进行编码;在解码过程中添加车辆自身运动学模型得到动态可行的预测轨迹。对公开数据集NGSIM进行实验分析,结果表明:模型对车辆轨迹预测误差在0.5 m以内;通过对轨迹预测的ADE与FDE结果分析,在预测未来5 s轨迹的情况下,相比于其他方法,ADE降低了30.7%,FDE降低了32.5%,验证了模型和算法的有效性。To address the low prediction accuracy in long-term trajectory prediction due to changing environments and interactive influences among vehicles,this paper proposes a multi-agent trajectory prediction model based on the Multiple Futures Predictor(MFP).First,the Symmetric Exponential Moving Average method is employed to remove outliers and smooth the trajectories.Then,the model utilizes Graph Convolutional Neural Network(GCN)for extracting interactive features between historical trajectories and future agents,encoding the interaction features.Finally,during the decoding process,the vehicle’s own kinematic model is incorporated to generate dynamically feasible predicted trajectories.Experimental analysis is conducted on the publicly available NGSIM dataset.Our results demonstrate the model achieves trajectory prediction errors within 0.5 m.Compared to the results of other methods,the proposed model reduces ADE(Average Displacement Error)by 30.7%and FDE(Final Displacement Error)by 32.5%when predicting trajectories within 5 seconds,validating the effectiveness of the model and algorithm.

关 键 词：自动驾驶 车辆轨迹预测 图神经网络 特征提取 MFP模型 

分 类 号：U491.23[交通运输工程—交通运输规划与管理]