One-Handed Gesture Recognition Based on Improved ST-GCN and CA Attention Mechanism

The article takes one-handed gesture video as the research object, and mainly focuses on the recognition process and method of “isolated words” and static gesture mixtures. Firstly, the algorithm of extracting the key point positions of frame pictures is proposed by grouping rules of frame sequences of different gestures. This algorithm solves the problem of invalid gesture images affecting the recognition due to the redundancy of gesture images when extracting gestures from traditional gesture videos. Finally, a new network (ST-GCN-CA) is proposed, which improves the traditional spatio-temporal graph convolutional network (ST-GCN) and adds the CA attention mechanism. The method and network in this paper are applied to the Chinese isolated word sign language dataset (CSL-500) single-handed video, and the experimental results show that the recognition accuracies in the isolated word gesture video and the hybrid video reach 94.3% and 94.1%, respectively, which improves the recognition accuracies in comparison with other methods in the public domain.