Wireless Human Activity Recognition: A Survey of Recent Advances and a Novel Zero-Shot LLM-Based Inference Framework

Abstract

With the rapid advancement of wireless sensing technologies, wireless human activity recognition (HAR) has emerged as a prominent, low-cost, and non-intrusive research area over the past decade. Despite these advances, wireless HAR remains fundamentally challenged by device heterogeneity, human body diversity, and environmental variability, factors that typically require large-scale labeled radio frequency data for model training. This paper first presents a systematic review of representative techniques and publicly available datasets, followed by a comprehensive analysis of recent progress and persistent limitations of HAR. Building on these insights, we propose a novel zero-shot inference framework based on large language models that directly infers human activities from channel state information sequences. By eliminating the need for activity labels from the target dataset, the framework enables faster adaptation to device heterogeneity, human body diversity, and environmental variability. Specifically, we integrate activity model knowledge, which captures the physical interactions between human motion and wireless signal propagation, and global temporal knowledge, which encodes coarse-grained temporal similarity patterns, into structured chain-of-thought prompts. This design empowers the framework to perform semantic and physical reasoning over raw wireless observations. Extensive experimental results demonstrate that the proposed approach achieves robust zero-shot performance, attaining an accuracy of 0.93 and an F1-score of 0.94. This performance significantly outperforms existing zero-shot state-of-the-arts.