基于GCN和Transformer时空信息融合的行人轨迹预测  

Pedestrian trajectory prediction based on spatio-temporal information fusion using GCN and Transformer

作　　者：柳军杰 蔡英[1] 范艳芳[1] 赵放 LIU Junjie;CAI Ying;FAN Yanfang;ZHAO Fang(College of Computer Science,Beijing Information Science&Technology University,Beijing 102206,China)

机构地区：[1]北京信息科技大学计算机学院,北京102206

出　　处：《北京信息科技大学学报（自然科学版）》2024年第6期1-8,共8页Journal of Beijing Information Science and Technology University（Science and Technology Edition）

基　　金：国家自然科学基金项目(61672106);北京市自然科学基金项目(L192023)。

摘　　要：针对行人轨迹的多态性导致模型预测精度下降的问题,提出了基于图卷积网络(graph convolution network,GCN)和Transformer时空信息融合的行人轨迹预测模型。采用Transformer提取行人运动行为特征;构建时空图,并根据结伴行人轨迹相似的特点,设计基于轨迹相似度的GCN,来优化时空图的卷积以提取行人间的空间交互特征;针对时空图之间的关联性,采用Transformer编码不同历史时刻的空间交互特征,深度挖掘行人轨迹的时空交互特征。融合行人运动行为特征、空间交互特征和时空交互特征实现基于时空信息融合的行人轨迹预测。在ETH-UCY和SDD公开数据集上的实验结果验证了所设计模型的性能和有效性。Regarding the decrease of model prediction accuracy caused by the polymorphism of pedestrian trajectories,a pedestrian trajectory prediction model based on spatio-temporal information fusion using the graph convolution network(GCN)and Transformer was proposed.Transformer was used to extract the movement behavioral features of pedestrains.A spatio-temporal graph was constructed,and a GCN based on trajectory similarity was designed according to the similarity of the trajectories of the partnered pedestrians,to optimize the spatio-temporal graph convolution to extract the spatial interaction features among pedestrians.Given the correlation between spatio-temporal graphs,Transformer was used to encode the spatial interaction features at different historical moments,in order to deeply explore the spatio-temporal interaction features of pedestrian trajectories.Pedestrian trajectory prediction based on spatio-temporal information fusion was achieved by integrating pedestrian movement behavior features,spatial interaction features,and spatio-temporal interaction features.Experimental results on public datasets like ETH-UCY and SDD validate the performance and effectiveness of the proposed model.

关 键 词：行人轨迹预测 TRANSFORMER 图卷积网络 轨迹相似度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]