WiFi-based activity recognition using activity filter and enhanced correlation with deep learning

Device-free WiFi sensing utilizing channel state information (CSI) is attractive for human activity recognition (HAR). However, several challenging problems are yet to be resolved, e.g., difficulty in extracting proper features from input signals, susceptibility to the phase shift of CSI and difficulty in identifying similar behaviors (e.g., lying and standing). In this paper, we aim to tackle these problems by proposing a novel scheme for CSI-based HAR that uses activity filter-based deep learning network (HAR-AF-DLN) with enhanced correlation features. We first develop a novel CSI compensation and enhancement (CCE) method to compensate for the timing offset between the WiFi transmitter and receiver, enhance activity-related signals and reduce the dimension of inputs to DLN. Then, we design a novel activity filter (AF) to differentiate similar activities (e.g., standing and lying) based on the enhanced CSI correlation features obtained from CCE. Extensive simulation results demonstrate that our proposed HAR-AF-DLN scheme outperforms state-of-the-art methods with significantly improved recognition accuracy (especially for similar activities) and notably reduced training time.