Variable Viewpoint Gesture Recognition Based on a Hybrid Graph Neural Network

The variations in camera view and hand spatial pose are the main reasons for the low accuracy and poor robustness of gesture recognition systems. In order to achieve accurate and stable gesture recognition with variable viewpoint, this article carries out research based on 3-D non-Euclidean vector graph features and graph neural networks. First, the 3-D information of hand joints is collected to construct a graph-structured gesture feature dataset, and a joint-based 3-D non-Euclidean vector graph method is proposed to solve the problem that similar gesture features are overly sensitive to spatial position and angle changes. Then, a Multi-Head graph attention network is designed and combined with graph convolutional neural network to explore the optimal hybrid graph neural network gesture recognition model. The experimental results show that, on the dataset processed by the joint-based 3-D non-Euclidean vector graph method, the training, testing, and validation accuracies of the optimal model reach 97.07%, 96.95%, and 87.06%, which are increased by 18.46%, 18.88%, and 44.23% compared to the original dataset, respectively. In conclusion, the method in this article is not only more robust to the variable viewpoint gesture recognition problem, but also has the advantages of low computational resource requirement and high real-time performance.