DARFNet: Dual-Branch Adaptive Representation Fusion Network for Cross-Modal Behavior Recognition Using WiFi and RFID Signals

In behavior recognition tasks, traditional visual methods suffer from privacy leakage, environmental sensitivity, and high costs, while existing WiFi or RFID single-modal methods are limited by multipath effects and low sampling rates. To address the challenges of heterogeneous modality fusion in wireless signal behavior recognition, this paper proposes the Dual-Branch Adaptive Representation Fusion Network (DARFNet). The network adopts a dual-stream Transformer architecture, enhances the Transformer module (ENTR), and constructs a Weighted Sum Nonlinear Fusion Module (WS-NTM) to achieve cross-modal adaptive weighting. Experiments using the XRF55 dataset show that the fusion model achieves an accuracy rate of 88.41 %, significantly outperforming existing methods. Ablation experiments indicate that the WS-NTM achieves an average gain of 2.08 % over traditional concatenation fusion, while the ENTR module contributes a $\mathbf{1. 6 8 \%}$ improvement in accuracy. This solution provides reliable technical support for medical monitoring and smart home scenarios.