.png){kind=icon}
Return
Hierarchical adaptive multi-scale hypergraph attention convolution network for skeleton-based action recognition
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asoc.2025.112855.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The Graph Convolutional Networks (GCN) have been widely used for skeleton-based human action recognition (HAR). However, the inherent graph topology design strategy and equally treated the spatial and temporal information in feature aggregation are two main problems existed in GCN-based skeleton HAR. To effectively address those issues, a hierarchical adaptive multi-scale hypergraph attention convolution network (HAMHGNet) is proposed in our work. Firstly, a hierarchical adaptive clustering partition module is formulated to construct dynamic graph topology for inferring the joint- to parts -to group interaction according their movements in different actions. Then, a multi-scale hypergraph attention convolution module is designed to relax the restriction of the fixed topology, maintaining the inherent spatial characteristic of the skeleton joints. Finally, a temporal segmentation attention constrained encoding module is constructed to model the relationship between different joints in several consecutive frames. Experimental results on two benchmark datasets, i.e. NTU-RGB+D 60 (93.1 % on X-View and 96.7 % on X-Sub) and NTU-RGB+D 120 (90.1 % on X-Sub and 90.8 % on X-Set), validate the proposed model can achieve the state-of-the-art performance for skeleton-based action recognition.
Keywords:
Action recognition
Multi-scale hypergraph attention convolution
Hierarchical adaptive clustering
Dynamic graph topology
Journal
IF:
6.6
Papers:
1.4W
Citations:
4.8W
Organization
No organization information available
Cited Papers
ShareSaveHierarchical graph attention network with pseudo-metapath for skeleton-based action recognition.png)
NEUROCOMPUTING
NEUROCOMPUTINGIF6.5
PREAI
ShareSave
ShareSaveShareSaveShareSaveShareSave
View more