This paper proposes the Dual-Stream Gated Fusion Transformer (DS-GFT) for WiFi Channel State Information (CSI)-based human activity recognition, addressing limitations in capturing both transient and stable temporal patterns. WiFi CSI provides cost-effective, non-invasive, and privacy-preserving sensing, yet existing methods struggle with feature balance. DS-GFT integrates parallel dynamic and static streams in which the dynamic stream detects instantaneous signal variations while the static stream extracts persistent features. A gated fusion module optimally combines these complementary representations, enhanced by a lightweight Transformer for temporal modeling. Evaluations on benchmark datasets demonstrate state-of-the-art 98.94% accuracy, validating superior robustness and generalization. This work establishes an effective architecture for temporal feature learning in wireless sensing systems, showing practical potential for smart home, healthcare, and security applications.