基于自适应时空图卷积的骨骼行为识别方法OA
Adaptive spatio-temporal graph convolutional network for skeleton-based action recognition
大多数骨骼行为识别方法通过预定义的骨骼拓扑结构来建模空间关系.然而,该方法所依赖的预定义拓扑存在本质缺陷,其静态性与共享性限制了模型特征提取的灵活性.为解决该问题,提出了一种基于自适应时空图卷积网络(adaptive spatio-temporal graph convolutional network,AST-GCN)的骨骼行为识别方法.该网络包含自适应空间图卷积(adaptive spatial graph convolution,AS-GC)与自适应时间图卷积(adaptive temporal graph convolution,AT-GC)两个核心模块.自适应空间图卷积通过多个可学习的邻接矩阵替代固定拓扑结构,自适应地提取动作的空间特征.自适应时间图卷积则在时间维度进行多层次建模,通过可学习邻接矩阵实现对不同时间段动态信息的细化学习.AST-GCN在大规模数据集NTU-RGB+D 60上的准确率为92.9％(X-Sub)和96.9％(X-View),在NTU-RGB+D 120上的准确率为89.7％(X-Sub)和91.0％(X-Set),在性能表现上优于现有的主流方法,验证了模型的有效性.
Most skeleton-based action recognition methods model spatial relationships using a predefined skeletal topology.However,this pre-defined topology suffers from inherent limitations,as its static and shared nature restricts the flexibility of feature extraction.To address this issue,this paper proposed an AST-GCN for skeleton-based action recognition.The network consists of an AS-GC and an AT-GC.The AS-GC replaced the fixed topology with multiple learnable adjacency matrices to adaptively extract spatial features of poses.The AT-GC modeled the temporal dimension at multiple levels,enabling detailed learning of dynamic information across different time segments through learnable adjacency matrices.AST-GCN achieved an accuracy of 92.9％(X-Sub)and 96.9％(X-View)on the large-scale dataset NTU-RGB+D 60,and 89.7％(X-Sub)and 91.0％(X-Set)on NTU-RGB+D 120.Its performance surpasses that of existing mainstream methods,demonstrating the ef-fectiveness of the model.
徐洪;吕凯;袁亮
新疆大学 软件学院,乌鲁木齐 830046新疆大学 机械学院,乌鲁木齐 830046新疆大学 软件学院,乌鲁木齐 830046||新疆大学 机械学院,乌鲁木齐 830046||上海交通大学南加州大学文化创意产业学院,上海 200240
信息技术与安全科学
行为识别图卷积网络自适应时空拓扑
action recognitiongraph convolutional networkadaptivespatio-temporal topology
《计算机应用研究》 2026 (5)
1594-1600,7
国家自然科学基金资助项目(62501517,52275003)中央高校基本科研业务费专项资金资助项目(buctrc202105)
评论