Edge-Enabled Human Activity Recognition Using Hybrid Deep Learning and Multi-Sensor Wi-Fi CSI Arrays

Recent advances in wireless sensing have positioned Wi-Fi Channel State Information (CSI) as a promising modality for human activity recognition (HAR), offering a cost-effective and non-intrusive alternative to traditional vision- and wearable-based systems. However, existing CSI-based HAR approaches suffer from high signal volatility, limited feature representation capacity, and poor adaptability in occluded or dynamic environments—particularly in edge or low-altitude IoT deployment scenarios. To overcome these challenges, we propose an edge-enabled HAR system based on a 2×2 array of Wi-Fi CSI sensors, which synchronously capture multi-channel signals to enhance sensing robustness and spatial diversity. Sensor data is processed locally at the edge via a compact, on-device deep learning framework that integrates convolutional, recurrent, and attention mechanisms. This hybrid architecture is specifically optimized for extracting both spatial and temporal motion features from noisy and rich scattering environments, improving recognition accuracy while reducing transmission and computation overhead. Experiments under both line-of-sight and non-line-of-sight conditions demonstrate the proposed system achieves recognition accuracies of 99% and 97%, respectively, outperforming conventional single-receiver models and demonstrating strong robustness in complex environments. By tightly coupling CSI-based wireless sensing with edge AI, our approach represents a robust and scalable solution for next-generation HAR in complex, low-latency, and privacy-sensitive IoT environments.