DAMNet: A Lightweight Dual-Modal Network for Wi-Fi-Based Human Activity Recognition

Wi-Fi signals have played an important role in human activity recognition (HAR) due to widespread usage and noninvasive nature. Whereas the single-modal feature of channel state information (CSI) is usually used but fails to fully provide the time-frequency characteristics of the Wi-Fi signal, which affects the classification accuracy of HAR. Therefore, this article proposes a lightweight dual-modal network DAMNet, which can extract and fuse the characteristics of the wireless signal in the time domain and frequency domain. The DAMNet enhances the complementarity between different modals to improve the representation ability of the model by constructing an inverted residual coordinate attention block (IRCAB) and local convolutional aggregation structures. Hence, it can more effectively capture the time-frequency characteristics of Wi-Fi signals and enhance the ability to recognize different activities. The experimental results on the ACSID dataset show that the accuracy of the proposed model is 98.10%, which is significantly better than those of the other most advanced models. Meanwhile, the generalization validation on the WiAR and Glasgow datasets is conducted, whose accuracy reaches 95.91% and 95.15%, respectively. It is proved that the method proposed in this article has excellent generalization performance, with important practical value and promotion significance.