无线电工程2025，Vol.55Issue(6)：1152-1160,9.DOI:10.3969/j.issn.1003-3106.2025.06.003
跨域图卷积网络在基于骨架动作识别中的应用
Cross-domain Graph Convolution Network for Skeleton-based Action Recognition
摘要
Abstract
In skeleton-based action recognition,Graph Convolutional Neural Network(GCN)has become a key technology due to their effectiveness in processing non-Euclidean data.However,existing GCN-based methods often restricted to single-temporal neighborhood information,and overlook the dynamic evolution and complex spatial-temporal dependencies in actions,thus limiting their performance in capturing fine-grained variations and long-term dependencies within continuous action sequences.To address these limitations,an innovative graph convolutional architecture is introduced.By cross-domain neighborhood message passing and aggregation,coupled with the integration of node state transitions over multiple timesteps,the model's representational power across varying temporal and spatial scales is boosted.Additionally,to further improve the discriminative capability for complex action recognition,a bidirectional multi-stream skeleton information fusion framework is designed to analyze human skeletal data from multiple perspectives,sensing subtle movements of joints and bones,and effectively filtering out disturbances from static frames.This is critical in discerning similar motion patterns.Experiments are conducted on three representative benchmark datasets NTU RGB+D 60,NTU RGB+D 120,and UAV Human to validate the effectiveness of the proposed method.关键词
图卷积神经网络/基于骨架的行为识别/邻域信息/时空关系图Key words
GCN/skeleton-based action recognition/neighborhood information/spatial-temporal graph分类
计算机与自动化引用本文复制引用
王欣霏,龚勋,吕金荣..跨域图卷积网络在基于骨架动作识别中的应用[J].无线电工程,2025,55(6):1152-1160,9.基金项目
国家自然科学基金(62376231) （62376231）
四川省自然科学基金(24NSFC1070,2023NSFC1616) National Natural Science Foundation of China(62376231) （24NSFC1070,2023NSFC1616）
Sichuan Provincial Natural Science Foundation of China(24NSFC1070,2023NSFC1616) （24NSFC1070,2023NSFC1616）
