Enhanced Wi-Fi Sensing: Leveraging Phase and Amplitude of CSI for Superior Accuracy

Human activity recognition (HAR) in indoor environments using Wi-Fi channel state information (CSI) remains a topic of significant interest and has seen rapid development in recent years. This field offers great potential due to the widespread availability of Wi-Fi signals and their cost-effectiveness, with applications in areas such as elderly care systems, context-aware environments, and security monitoring. However, Wi-Fi-based HAR faces considerable challenges in maintaining consistent performance across various environments and individuals, primarily due to the inherent variability of Wi-Fi signals. Addressing this challenge requires training models on large and diverse datasets to ensure robustness and generalization across different environments and conditions. Furthermore, most existing research focuses on recognizing human gestures by extracting a single feature from Wi-Fi signals over time sequences, while neglecting multi-channel signal information and underutilizing other valuable features. To overcome this limitation, we propose a model named PA-CSI that leverages both amplitude and phase features from Wi-Fi signals, incorporating attention mechanisms across both temporal and channel dimensions, along with multi-scale convolutional neural networks (CNNs). The PA-CSI model demonstrates competitive accuracies across datasets: 99.9% on StanWiFi, 98% on MultiEnv, and 99.9% on a self-constructed dataset. The source code is available at https://github.com/thai-duy-quy/PA-CSI-HAR.