Research on Human Action Recognition Based on WiFi CSI Using Improved Parallel Neural Network

Device-free human activity recognition (HAR) using WiFi Channel State Information (CSI) offers a low-cost and privacy-preserving sensing solution. However, existing deep learning approaches often rely on generic architectures that underutilize multi-domain physical features and incur prohibitive computational overhead for edge devices. In this paper, we propose a lightweight, task-oriented parallel neural network tailored for WiFi sensing. Specifically, high-resolution Doppler Frequency Shift (DFS) features are explicitly extracted to capture fine-grained human kinematics. These are jointly modeled alongside temporal and spatial dynamics using a three-branch architecture comprising attention-enhanced bidirectional GRUs and a channel-attention CNN module. Furthermore, an adaptive attention-gated mechanism is introduced for optimal multidomain feature fusion. Extensive experiments on the public UT-HAR dataset yield a state-of-the-art accuracy of 98.49%. Additional evaluations on the highly challenging, cross-domain Widar dataset demonstrate an accuracy of 83.33%, substantially outperforming classical and heavy-weight vision baselines (e.g., ResNet18 and ViT). Requiring only 4.92 M parameters and 98.72 M FLOPs, the proposed framework achieves a superior trade-off between recognition precision and computational efficiency, validating its immense potential for practical, resource-constrained WiFi sensing applications.