Human activity sensing based on WiFi Channel State Information has emerged as a prominent research focus in recent years. To address the challenges of ineffective activity data extraction and the reliance on handcrafted features in existing methods, this paper proposes a WiFi-based human activity extraction and sensing method that integrates grouped multi-scale attention. First, amplitude and phase features most sensitive to human activity are extracted using antenna selection and conjugate multiplication, followed by wavelet-based denoising for preprocessing. Next, we compute sliding window factors and activity correlation factors through the variance of joint amplitude and phase features, and apply an empirical threshold to extract valid activity sequences. Finally, we design an end-to-end deep learning model (D-GMAFi) based on a grouped multi-scale attention mechanism to capture long-range temporal dependencies, multi-scale spatial information, and dynamically emphasize critical features across different dimensions. Extensive simulation experiments conducted on the self-collected WiACT dataset demonstrate that D-GMAFi achieves average recognition accuracies of 94.63% and 91.06% under single-person single-scene and multi-person multi-scene settings, respectively. The proposed method outperforms the baseline models overall and maintains stable recognition performance across different users and scenarios, indicating a certain degree of robustness and adaptability to application-oriented settings. These results suggest that the proposed approach exhibits strong potential for WiFi-based human activity recognition tasks and can provide a useful reference for the practical application of WiFi sensing technologies in real-world scenarios.