WiGNN: WiFi-Based Cross-Domain Gesture Recognition Inspired by Dynamic Topology Structure

Gesture recognition based on WiFi signals has achieved significant progress with the advent of deep learning. The channel state information (CSI) carried by the WiFi signal, is commonly used in deep learning-based models to extract features of human activities. However, when a user's location, orientation, or other gesture-independent information changes, the recognition accuracy of the model generally degrades significantly, introducing the challenge in cross-domain gesture recognition. This article reviews recent efforts in WiFi-based cross domain gesture recognition, which are faced with the challenge of accuracy and generalization. Observing that the movement of a user introduces variations in CSI simultaneously at multiple WiFi receivers, we capture the spatio- temporal relationship of the signals received at different spots with the graph model. We propose the causal multi-scale temporal-frequency feature fusion layer, which realizes extraction of temporal features with a casual convolution, followed by a multi-scale frequency extractor dealing with the rich frequency components in the CSI data. Embedding the temporal-frequency feature in the graph node, a graph neural network adaptively aggregates features with respect to different gestures. Moreover, to improve the robustness of possible signal obstructions caused by human orientations, a data augmentation scheme is proposed based on the spatial relationship between the receivers and the user. Our model achieves the average accuracy of 94.0 percent in cross-domain tasks on the Widar3.0 dataset, demonstrating the superiority of WiGNN.