WiSwin: A Lightweight WiFi Gesture Recognition Framework with Swin Transformer

Deep learning based human activity recognition with WiFi signals has become one of the most promising applications of WiFi technology. However, gesture recognition with high inand cross-domain accuracy remains a challenging problem. In this study, we propose a novel lightweight gesture recognition framework with WiFi signals, named WiSwin, which consists of three modules: a data transformation module that converts estimated channel state information (CSI) from WiFi signals into an image sequence to extract environment-independent features, a Swin Transformer module and a BiLSTM module that extract, respectively, spatial and temporal features of the converted image sequence. Extensive numerical results demonstrate that the proposed low-complexity WiSwin framework achieves superior recognition accuracy to the state-of-the-art methods, with the highest in-domain accuracy of $99.8 \%$ while maintaining competitive cross-domain accuracy.