Neural Approaches to Wi-Fi CSI-Based Human Activity Recognition: Architectures and Generalization

WiFi-based passive sensing technology has become a key enabler in the rapidly advancing field of wireless communications, supporting a wide range of applications such as human activity recognition (HAR), gesture recognition, and indoor localization. Traditional WiFi sensing methods primarily rely on the received signal strength indicator (RSSI), which provides coarse-grained information and suffers from limited reliability. With the emergence of fine-grained channel state information (CSI), passive sensing using WiFi has achieved significant improvements in accuracy, robustness, and versatility. Recent advances in deep learning (DL) have greatly accelerated the development of WiFi CSI-based sensing, owing to its powerful capabilities in automatic feature extraction and complex pattern modeling. This survey provides a comprehensive and up-to-date overview of the latest developments in DL-based WiFi sensing techniques for HAR. The discussion begins with an overview of signal propagation characteristics and CSI acquisition methods, followed by a classification of sensing tasks based on data modalities and application scenarios. We then present a detailed review on commonly adopted neural network architectures, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), attention-based models, and Transformer variants. Furthermore, we examine recent methodological advancements such as domain adaptation, continual learning, and self-supervised learning. A comparative analysis highlights the strengths, limitations, and practical performance of various approaches. Finally, we discuss key challenges in the field, including cross-domain generalization, real-time processing, privacy concerns, and multi-modal fusion. We outline potential research directions to support the development of robust, scalable, and deployable WiFi sensing systems in real-world environments.