WiFi-Based Indoor Human Activity Sensing: A Selective Sensing Strategy and a Multilevel Feature Fusion Approach
Y. Zhang, Gongpu Wang, Heng Liu, Wei Gong, Feifei Gao
Abstract
Utilizing communication signals for indoor human activity recognition (HAS) is an important component of integrated sensing and communication (ISAC). The current majority HAS solutions adopt a single sensing strategy and only work in a simple environment. In this paper, we propose a new HAS method named WiSMLF that can flexibly select multiple sensing strategies and then use multi-level feature fusion for sensing. We first use the high frequency energy (HFE) method to categorize human activities into two types: static activities (SAs) and moving activities (MAs). Subsequently, for SAs, we adopt a joint localization and activity recognition sensing strategy, and use a multi-level feature fusion network based on visual geometry group (VGG). For MAs, we adopt a joint activity recognition and moving distance estimation sensing strategy, and use a multi-level feature fusion network based on long short-term memory (LSTM). The experimental results show that WiSMLF outperforms the existing methods especially in complex environments, and can obtain 92% higher accuracy in location, activity recognition, and distance estimation.