Multi-attention graph convolutional network for skeleton-based action recognition

Action recognition is a challenging task in computer vision, playing a pivotal role in understanding human behavior. Graph convolutional neural networks naturally excel in extracting spatiotemporal features from human skeleton sequences, finding extensive applications in skeleton-based action recognition. Skeleton sequences contain rich information across the dimensions of time, space, and channels. However, current methods lack a balanced design for information extraction in these three dimensions, with many models being overconfident during the training process. In this paper, we present the Multi-Attention Graph Convolutional Network (MA-GCN), utilizing attention modules designed across these three dimensions for the purpose of feature extraction and fusion. We construct channel and spatial attention modules separately by using different pooling methods, establish temporal attention module through a combination of local context and global context. During the training process, we utilize label smoothing regularization to constrain the model, enhancing its generalization ability. We conducted extensive experiments using the NW-UCLA, NTU RGB+D, and NTU RGB+D 120 datasets. Our approach demonstrated its effectiveness by surpassing the baseline method by 0.5%, 0.3%, and 0.4% on the two large-scale datasets.