Enhanced Adjacency Matrix-Based Lightweight Graph Convolution Network for Action Recognition

Graph convolutional networks (GCNs), which extend neural (CNNs) to non-Euclidean structures, have been utilized promote skeleton-based human action recognition research and made substantial progress in doing so. However, there are still some challenges the construction of models based on GCNs. In this paper, we propose an enhanced adjacency matrix-based graph network with a combinatorial attention mechanism (CA-EAMGCN) for recognition. Firstly, matrix is constructed expand model’s perceptive field global node features. Secondly, feature selection fusion module (FSFM) designed provide optimal ratio multiple input features model. Finally, devised. Specifically, our spatial-temporal (ST) limb (LAM) integrated into multi-input branch mainstream proposed model, respectively. Extensive experiments three large-scale datasets, namely NTU RGB+D 60, 120 UAV-Human show that model takes account both requirements light weight accuracy. This demonstrates effectiveness method.