WiDeFus: A Wi-Fi-Based Lightweight Human Activity Recognition via CSI Component Decomposition and Adaptive Feature Fusion

WiFi channel state information (CSI)-based human activity recognition (HAR) approaches face some fundamental limitations, such as high computational cost of the model due to irrelevant signal components obscure human-related features, position-sensitive representations cause activity misclassification under spatial variance, and environmental heterogeneity induces domain shifts that degrade generalization. To overcome these challenges, we propose a lightweight HAR approach based on WiFi CSI component decomposition and triple feature fusion (WiDeFus). Specifically, WiDeFus first isolates human-related components via quantum-inspired sparse decomposition, and leverage Hermite-Gaussian bases with group-element sparsity constraints to eliminate dynamic interference and noises. WiDeFus then introduces a triple-feature adaptive fusion network to achieve dynamic frequency-domain selection, automatically extract temporal features, and perform environment-robust calibration. These purified features are processed by a dendrite net (DD) that replaces nonlinear activations with multiplicative interactions for efficient inference. Experimental results show that WiDeFus is a lightweight HAR approach with high recognition accuracy and satisfactory cross-domain generalization performance.