A Generalizable Prompt-Based Prototypical Framework for CSI-based Few-Shot and Cross-Domain Activity Recognition

Wireless sensing systems for human activity recognition (HAR) have made great strides in recent years. However, current HAR models face challenges in generalization due to a limited number of training samples (i.e., few-shot problem) and pattern differences of the collected channel state information (CSI) data across diverse domains (i.e., cross-domain problem). To address the above problems, in this paper, we propose a prompt-based prototypical framework called Wi-Prompt. Our Wi-Prompt framework consists of the following three modules: Prompt generation module, prototypical representation module, and activity recognition module, which works as follows. Prompt generation module is developed to extract prior knowledge from samples in the source domains, providing insightful guidance for establishing class prototypes in the target domains. Prototypical representation module effectively captures the most representative prototype vector of each class with a temporal convolutional network (TCN)-attention model. Activity recognition module determines the activity class of new sample by comparing the Euclidean distance between its corresponding prototype vector and the prototype vector of each class. The greatest advantage of Wi-Prompt is its utilization of prompt-based prototype representation, which eliminates the need for prior domain-specific knowledge about the original CSI samples, making it highly adaptable to a wide variety of CSI datasets collected across different domains. Extensive experiment results based on real-world traces show that our proposed Wi-Prompt outperforms state-of-the-art models under various cross-domain scenarios.