An Efficient Human Activity Recognition System Using WiFi Channel State Information

Insufficient recognition precision and high complexity are two main challenges of human activity recognition using WiFi channel state information (CSI), which has attracted more attention due to its low cost and easy realization. To address these challenges, we propose a novel framework based on Gramian angular fields (GAFs). This framework includes two transformation methods, Gramian angular sum field (GASF) and Gramian angular difference field (GADF), which effectively extract information from CSI and convert it into a CSI-GAF image. Subsequently, a convolutional neural network (CNN) is designed to analyze these images and obtain activity information. By incorporating a transformation module that preserves and expands the original CSI information, the proposed framework utilizes the powerful feature extraction capabilities of the CNN in image processing. Test results on public CSI datasets (Wiar, SAR, and Widar3.0) demonstrate that the recognition accuracy based on the GADF outperforms that of GASF, reaching 99.4% and 99.0%, respectively, when the CNN has only four convolutional layers. Moreover, the proposed framework exhibits low complexity, which outperforms three classical models (ResNet, VGG19, and ShufflenetV2) in terms of both parameters and required floating-point computations.