Lightweight CSI-Based Human Activity Recognition with Cross-Environment Robustness

This paper proposes a lightweight human activity recognition (HAR) method based on Wi-Fi channel state information (CSI) sensing. The proposed architecture, termed DBA-HAR, incorporates depthwise separable convolutions, multi-head attention mechanisms, and a novel Dual-Branch Fusion Block (DBFB) to efficiently extract spatiotemporal features while significantly reducing computational complexity. A Gaussian encoding module is further introduced to enhance temporal sensitivity. Compared to existing models such as MLP, CNN-1D, CNN-2D, and THAT, DBA-HAR achieves superior performance with substantially fewer parameters and lower computational overhead (FLOPs). Extensive experiments conducted on a publicly available cross-environment CSI dataset demonstrate that DBA-HAR attains a recognition accuracy of 94.92 %.The model shows strong robustness in recognizing static and low-dynamic behaviors, and achieves notable improvements in overall efficiency. However, confusion matrix analysis indicates limitations in recognizing transitional actions, suggesting opportunities for future enhancement through improved temporal modeling. These results highlight the potential of DBA-HAR for deployment in intelligent healthcare, smart home systems, and ambient sensing applications.