Spatial adaptive graph convolutional network for skeleton-based action recognition

• DOI: 10.1007/s10489-022-04442-y
• Publication Date: 2023-01-13T05:04:15Z
• Authors: Qilin Zhu; Hongmin Deng
• Keywords: RGB color model; Adjacency matrix

Abstract In recent years, great achievements have been made in graph convolutional network (GCN) for non-Euclidean spatial data feature extraction, especially the skeleton-based extraction. However, fixed structure determined by adjacency matrix usually causes problems such as weak modeling ability, unsatisfactory generalization performance, excessively large number of model parameters, and so on. this paper, a spatially adaptive residual (SARGCN) is proposed action recognition based on skeleton Firstly, uniform topology not required our graph. Secondly, learnable parameter added to GCN operation, which can enhance model’s capabilities extraction generalization, while reducing parameters. Therefore, compared with several existing models mentioned least parameters are used ensuring comparable accuracy. Finally, inspired ResNet architecture, connection introduced obtain higher accuracy at lower computational costs learning difficulties. Extensive experimental two large-scale datasets results validate effectiveness approach, namely NTU RGB+D 60 120.