A Lightweight WiFi-Based Human Activity Recognition Method for Unknown Domains

The unsupervised domain adaptation (UDA) strategy is widely used to solve cross-domain human activity recognition based on the WiFi channel state information (CSI). However, the UDA strategy requires collecting large unlabeled target domain data in advance, which is highly impractical for real-world deployments. Considering the strict unavailability of target domain data in practical single-transceiver scenarios, a lightweight recognition method based on the domain generalization (DG) strategy is proposed. To extend the spatiotemporal features of the activity samples, the raw amplitude and phase of the CSI data are integrated using the proposed Pseudo-Heatmap imaging algorithm. Based on the multi-source domain generalization strategy, a cross-domain recognition model, DGWiAR, is designed. DGWiAR consists of a data augmentation module, a teacher-student distillation module, a feature alignment module, and a classifier. The model divides the learning of domain-invariant features into two parallel parts: inter-domain and intra-domain. By embedding a deep residual network as the teacher network to guide the lightweight student network, the model can achieve a balance between the intra-domain invariant feature extraction ability of the student network and the computational cost. By using different source domains to imitate the unknown target domain, CORAL is utilized to align the feature distributions between the source and pseudo-target domains to learn inter-domain invariant features. Meanwhile, a constraint based on cosine distance is used to maintain the differences between inter-domain and intra-domain features to prevent the two types of features from crossing during the learning process. Additionally, considering the temporal characteristics of Pseudo-Heatmap images, DGWiAR introduces a data augmentation module based on a custom masking strategy to alleviate overfitting during training and enhance the robustness of the model to noise. Some available data has been used to train the model, then the trained lightweight student network is used to recognize activities from other unknown domains. The experimental results show that the accuracy of DGWiAR achieves 95.5%, 94.4%, 91.3%, and 93.0% across locations, orientations, users, and environments of a single link, respectively, and effectively reduces the complexity simultaneously.