全文下载:
202602021.pdf
文章编号:1672-6987(2026)02-0148-11;DOI:10. 16351/j. 1672-6987. 2026. 02. 021
杨 茜 a ,葛 艳 a*,杜军威 b ,陈 卓 b (青岛科技大学 a. 信息科学技术学院;b. 数据科学学院,山东 青岛 266061)
摘 要:针对行人轨迹预测任务中行人在空间维度的交互特征与时间维度上的时间模式难以 准确提取的问题,提出了 SocialReFSTGCN(social refined features spatiotemporal graph convolutional neural network)模型。在空间特征提取部分,提出了空间稀疏自注意力图卷积网络 (spatial sparse self-attention GCN,SS-GCN),利 用 空 间 稀 疏 自 注 意 力(spatial sparse selfattention,SSAttn)来突出行人间重要的空间交互关系并抑制冗余信息,从而使模型提取到的 空 间 特 征 更 加 准 确 与 精 炼 。 在 时 间 特 征 提 取 部 分 ,提 出 了 时 间 自 卷 积 模 块(time selfconvolution block,TSCBlock)代替原本在时间维度上的普通卷积,在提取丰富时间上下文的 同时使用自适应池化使空间信息有效保留。在预测阶段,使用优化的时间外推卷积神经网 络,增强沿时间维度的信息交互并进行轨迹预测。基于公开数据集 ETH 和 UCY 的实验结果 表明,所提出的方法在平均位移误差(ADE)和最终位移误差(FDE)指标上性能与同类模型相 比有良好提升。
关键词:行人轨迹预测;行人社会交互;时空图卷积网络;空间注意力机制;时间卷积
中图分类号:TP 181 文献标志码:A
引用格式:杨茜,葛艳,杜军威,等 . 基于稀疏自注意力时空图卷积的行人轨迹预测方法[J]. 青岛科技大学学报(自然科学版),2026,47(2):148-158.
YANG Qian, GE Yan, DU Junwei, et al. Pedestrian trajectory prediction method based on sparse self-attention spatio-temporal graph convolution[J]. Journal of Qingdao University of Science and Technology(Natural Science Edition),2026,47(2):148-158.
Pedestrian Trajectory Prediction Method Based on Sparse Self-Attention Spatio-Temporal Graph Convolution
YANG Qiana ,GE Yana ,DU Junweib ,CHEN Zhuob (a. College of Information Science and Technology;b. College of Data Science, Qingdao University of Science and Technology, Qingdao 266061,China)
Abstract:The task of pedestrian trajectory prediction poses challenges in accurately extracting spatial interaction features and capturing temporal patterns. Based on the above challenges, a new model named Social-ReFSTGCN (social refined features spatio-temporal graph convolutional neural network) based on spatio-temporal graphs is proposed. In the part of spatial feature extraction, the spatial sparse self-attention graph convolutional network (SS-GCN) is proposed, using spatial sparse self-attention to highlight important pedestrian interactions and suppress unnecessary redundant information, extracting features more accurately and refinedly. In the part of time feature extraction, the time self-convolutional block (TSCBlock) is proposed to replace ordinary convolution in the time dimension, extracting rich temporal context while effectively preserving spatial information by using auto pooling. During the prediction process, an optimized time extrapolation convolutional neural network (TXP-CNN) is employed to enhance the interaction of information along the temporal dimension. Experimental results based on public data sets ETH and UCY show that the proposed method has better performance in the mean average displacement error (ADE) and final displacement error (FDE) compared with simlar models using spatio-temporal graph convolution.
Key words:trajectory prediction; pedestrian social interaction; spatio-temporal graph convolu⁃ tional network; spatial sparse self-attention; time convolution
收稿日期:2025-03-09
基金项目:山东省自然科学基金项目(ZR2021MF092).
作者简介:杨 茜(1997—),女,硕士研究生 . * 通信联系人 .