摘要
自动驾驶领域下的智能体轨迹预测任务需要充分考虑到智能体与交通环境之间的关系。为了解决现有方法在异构特征交互层面的局限,提高预测精度,提出一种基于边更新与多头注意力交互融合Transformer的车辆轨迹预测方法EMATNet(edge-updating multi-head attention Transformer network)。该方法首先编码嵌入智能体与交通环境信息的历史时空信息;接着通过所提出的边更新与多头注意力交互融合Transformer两阶段式交互网络,引入对称位姿嵌入与车路关系交互,有效增强全局信息感知与时空关系捕捉能力;最终采用两阶段式优化解码确保预测结果的准确性与合理性。在Argoverse 1和Argoverse 2两个运动预测数据集验证模型有效性,并可视化分析预测结果。结果表明,EMATNet在minFDE、minADE、MR指标上均优于同类方法,能够胜任复杂交通环境车辆轨迹预测任务。
The task of vehicle trajectory prediction for autonomous driving needs to fully consider the relationship between the traffic agents and the environment.Addressing the limitations of existing approaches at the level of heterogeneous feature interaction and improving prediction accuracy,the paper proposed a vehicle trajectory prediction approach named EMATNet with edge updating and multi-head attention interactive fusion Transformer.Firstly,the approach encoded and embedded the histo-rical spatio-temporal information of the agents and the transportation environment.Then,the approach used the proposed two-stage interaction network of edge updating and multi-attention interaction fusion Transformer for feature interaction.The introduced symmetric positional embedding and vehicle-road relationship interaction could effectively enhance the global information perception and spatio-temporal relationship capturing capability.Finally,this approach used two-stage optimization decoding to ensure the accuracy and reasonableness of the prediction results.The proposed approach validated on Argoverse1 and Argoverse2 motion prediction datasets,and visualized and analyzed the prediction results.The results show that EMATNet outperforms similar approaches in the three performance metrics of minFDE,minADE and MR,and is capable for the task of vehicle trajectory prediction in complex traffic environments.
作者
孙颖
吴延勇
丁德锐
张建坤
Sun Ying;Wu Yanyong;Ding Derui;Zhang Jiankun(Business School,University of Shanghai for Science&Technology,Shanghai 200093,China;School of Optical-Electrical&Computer Engineering,University of Shanghai for Science&Technology,Shanghai 200093,China)
出处
《计算机应用研究》
北大核心
2025年第8期2348-2354,共7页
Application Research of Computers
基金
国家自然科学基金资助项目(62203306,62373251)。
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
vehicle trajectory prediction
deep learning
Transformer
multi-head attention mechanism
spatio-temporal feature integration
intelligent driving
