Wi-Cro: WiFi-Based Cross Domain Activity Recognition via Modified GAN

Abstract

WiFi-based Human Activity Recognition (HAR) plays a crucial role in achieving device-free Human-Machine Interaction (HMI). Due to the strong dependence of WiFi Channel State Information (CSI) on the target environment, current WiFi-based activity recognition methods often require extensive training data and struggle to adapt to new environments. To enhance the generalization ability of wireless sensing, we propose a cross-domain activity recognition system based on WiFi, WiCro. Wi-Cro utilizes Generative Adversarial Network (GAN) to transform source domain data into target domain data, supplemented by a small number of target domain samples to achieve cross-domain human activity recognition. We design a Subcarrier Dimension Reduction Algorithm (SDRA) to mitigate computational complexity of the system. This algorithm calculates the correlation between different CSI subcarriers and extracts principal components to reduce sample dimensions. Furthermore, we develop a data augmentation method based on GAN to transform source domain data into the target domain, addressing issues related to overfitting and domain adaptation due to insufficient samples. Finally, a metric learning approach is designed to further enhance the cross-domain sensing capability of the Wi-Cro system. Extensive experiments are performed using data collected in real-world environments and two public datasets. The results demonstrate that the Wi-Cro system consistently achieves high accuracy in human activity recognition across different datasets, proving its generalization abilities of cross-domain sensing.