无线电通信技术2025，Vol.51Issue(3)：576-587,12.DOI:10.3969/j.issn.1003-3114.2025.03.017
基于多尺度全局时空特征图的轨迹预测模型
Trajectory Prediction Model Based on Multi-scale Global Spatiotemporal Feature Maps
摘要
Abstract
A deep learning model based on a multi-scale global spatiotemporal feature map is proposed to address the challenge of modeling highly flexible pedestrians and vehicles in traffic environments.The model utilizes multi-scale graph encoding and decoding,incorporating hierarchical encoding of historical trajectory features for accurate prediction.Furthermore,to overcome the limited memo-ry capacity of traditional sequential models,the model introduces a self-attention mechanism based on spatiotemporal graphs,enhancing the memory capability of historical features and providing multiple prediction options for precise forecasting.Additionally,the model takes into account the global temporal features provided by pedestrians'intrinsic attributes,enriching the learnable features and strengthening temporal relationships.Experimental results on five benchmark datasets demonstrate that the proposed model achieves su-perior performance compared to existing models.The average displacement error was reduced by 17%,and the final displacement error was reduced by 52%.关键词
图神经网络/轨迹预测/时空特征/注意力机制Key words
graph neural network/trajectory prediction/spatiotemporal features/attention mechanism分类
信息技术与安全科学引用本文复制引用
韩新宇,李思照,徐火生,付小晶..基于多尺度全局时空特征图的轨迹预测模型[J].无线电通信技术,2025,51(3):576-587,12.基金项目
科技部重点研发计划(2022YFB4400703) （2022YFB4400703）
基础科研计划(JCKY2021604B002) （JCKY2021604B002）
中央高校基础科研业务费(3072024XX0601) Key Research and Development Program of the Ministry of Science and Technology(2022YFB4400703) （3072024XX0601）
Basic Scientific Research Program(JCKY2021604B002) （JCKY2021604B002）
Fundamental Research Funds for the Central Universities(3072024XX0601) （3072024XX0601）
