青岛大学学报（自然科学版）2025，Vol.38Issue(1)：14-23,10.DOI:10.3969/j.issn.1006-1037.2025.01.03
基于频域注意力时空图卷积网络的人体动作预测方法
Human Motion Prediction Method Based on Frequency Domain Attention Spatiotemporal Graph Convolutional Network
摘要
Abstract
To address the challenges in human motion prediction,such as the accumulation of prediction errors with increased prediction lengths,frame skipping of movements caused by the sudden increase of prediction errors,and difficulties in capturing long-range dependencies,a Spatiotemporal Graph Convolutional Network with Frequency Domain At-tention(SGFA)was proposed.The discrete cosine transform was employed to convert his-torical motion sequences into the frequency domain.The frame skipping issue was opti-mized,and long-range dependencies were captured through a multi-head self-attention mechanism,skeletal sequences were modeled using a graph convolutional network with an inverted residual structure to extract global features and generate action predictions,effec-tively reducing error accumulation.Human motion prediction experiments were conducted on the Human 3.6M dataset and the CMU MOCAP dataset for prediction horizons of 80 ms,160 ms,320 ms and 400 ms.The results show that SGFA further reduces the average joint angle error by 5%compared to the existing methods.关键词
人体运动预测/频域/注意力/图卷积网络Key words
human motion prediction/frequency domain/attention/graph convolutional network分类
计算机与自动化引用本文复制引用
童汗青,丁文文,李晴..基于频域注意力时空图卷积网络的人体动作预测方法[J].青岛大学学报（自然科学版）,2025,38(1):14-23,10.基金项目
国家自然科学基金(批准号:62171342)资助,安徽省高等学校自然科学基金(批准号:2024AH051682)资助. （批准号:62171342）
