Design of Human Action Recognition Method Based on Cross Attention and 2s-AGCN Model

This study explores the potential of human action recognition techniques for multi-disciplinary applications and the challenges of traditional methods. In recent years, although progress has been made in graph convolutional network-based approaches, they still fall short when dealing with highly dynamic and complex interactive actions. Therefore, the study proposes to introduce a cross-attention mechanism into the 2 s AGCN model to enhance the model recognition capability. The article provides an in-depth analysis of the graph convolution-based action recognition model and the attention mechanism, and describes the 2 s -AGCN model and its advantages in detail. Experiments are evaluated using the Kinetics-400 and NTU RGB+D 60 datasets, and the results show that the improved model has significantly improved recognition performance, and is especially more effective in handling complex interactions and fast action scenes. This result demonstrates the potential of deep learning in understanding human behaviour and lays the foundation for future developments.