Skeleton Action Recognition Based on Grouped Multi-scale Spatio-temporal Cross-Attention Graph Convolution

Action recognition has appeared as an important study area in computer vision, with graph convolutional networks (GCNs) are currently a mainstream research direction in skeletal action recognition. However, due to the weight bias between local nodes during the aggregation process, graph convolutional networks (GCNs) often dominate the aggregation process, failing to effectively establish strong correlations between unnaturally connected nodes. Attention mechanisms enhance key action nodes and frame information between distant joints. However, while attention weights are concentrated on globally salient regions, the relative importance of local features is automatically weakened, which leads to the neglect or elimination of neighborhood semantic features. To address this important issue, this research introduces a Grouped Multi-Scale Spatio-temporal Cross-Attention Graph Convolution (GMST-STCCA) to develop a network model that integrates joints, frames, and key local information crucial for the action. We proposed method outperforms most methods in terms of accuracy on the NTU-RGB+D and NTU-RGB+D 120 datasets.