融合复杂网络和记忆增强网络的轨迹预测技术  

Trajectory Prediction Technology Integrating Complex Network and Memory-Augmented Network

作　　者：赵高士 陈龙[1] 蔡英凤[1] 廉玉波 王海[2] 刘擎超[1] 滕成龙[1] Zhao Gaoshi;Chen Long;Cai Yingfeng;Lian Yubo;Wang Hai;Liu Qingchao;Teng Chenglong(Automotive Engineering Research Institute,Jiangsu University,Zhenjiang 212013;School of Automotive and Traffic Engineering,Jiangsu University,Zhenjiang 212013;Automotive Engineering Research Institute,BYD Automotive Industry Co.,Ltd.,Shenzhen 518118)

机构地区：[1]江苏大学汽车工程研究院,镇江212013 [2]江苏大学汽车与交通工程学院,镇江212013 [3]比亚迪汽车工业有限公司汽车工程研究院,深圳518118

出　　处：《汽车工程》2023年第9期1608-1616,1636,共10页Automotive Engineering

基　　金：国家重点研发计划(2022YFB2503302);国家自然科学基金(52225212,U20A20333,U20A20331,52072160);江苏省重点研发项目(BE2020083-3)资助。

摘　　要：周边目标轨迹预测是智能汽车决策规划的重要依据,现有基于多交通主体欧氏距离的建模方法无法有效描述多目标之间的复杂交互关系,制约其在实际动态交通场景中的适用性。本文创新地将复杂网络和记忆增强神经网络进行融合,构建了双层动态复杂网络模型,实现了高可靠性和可解释性的多模态轨迹预测。该模型使用高斯可变安全场计算风险权重,考虑了交通参与者的行驶状态参数、形状尺寸和智能体与道路之间的相互影响,真实准确地反映复杂环境中多交通主体间的交互关系;构建了一种由注意力机制和包含风险权重的社交池组成的复杂网络编码模块,实现了动态复杂场景中交通参与者与道路约束之间作用特征的全面有效提取;构建了以参考轨迹为条件的轨迹解码模块,实现了兼顾精度和可解释性的多模态轨迹输出。在公开数据集nuScenes上的验证结果表明,本文所提出的方法最小平均位移误差为1.37 m,最小最终位移误差为8.13 m,性能优异且具有较好的可解释性。The prediction of peripheral target trajectories is an important basis for intelligent vehicle decision-making and planning.Existing modeling methods based on multi traffic agent Euclidean distance cannot effectively describe the complex interaction relationship between multiple targets,which limits the applicability in practical dynamic traffic scenarios.In this paper,complex network and memory-augmented neural network are innovatively integrated to construct a double-layer dynamic complex network model to achieve high reliability and interpretability of multimodal trajectory prediction.This model uses a Gaussian variable safety field to calculate risk weights,taking into consideration of the driving state parameters,shape and size of traffic participants,as well as the interaction between intelligent agents and the road,truly and accurately reflecting the interaction relationship between multiple traffic agents in complex environments.A complex network-coding module composed of attention mechanism and social pool containing risk weights is constructed to realize comprehensive and effective extraction of interaction features between traffic participants and road constraints in dynamic and complex scenes.A trajectory-decoding module based on reference trajectories is constructed,realizing multimodal trajectory output that balances accuracy and interpretability.The validation results on the public dataset nuScenes show that the method proposed in this paper has a minimum average displacement error of 1.37 m and a minimum final displacement error of 8.13 m,with excellent performance and good interpretability.

关 键 词：智能汽车 轨迹预测 复杂网络 交互建模 

分 类 号：U463.6[机械工程—车辆工程]