APFNet: Amplitude-Phase Fusion Network for CSI-Based Action Recognition

Traditional Wi-Fi based action recognition models often utilize only amplitude information from CSI, with phase information simply discarded due to phase error problems. Such a design decision inevitably limits the information utilization, thus the overall performance. These models also strive at high recognition accuracy, invoking complicated model designs with high computational complexity, rendering them unsuitable for resource-constrained applications or swift migration. To address these issues, a novel lightweight action recognition model based on amplitude-phase fusion is proposed in this paper. To achieve maximum fusion benefit, the phase error is first corrected before information extraction. A lightweight multi-data fusion network termed APFNet is then designed and applied to fuse amplitude and phase information before extracting features for action recognition. Extensive experiments confirmed that comparing to schemes utilizing only amplitude information, APFNet can not only adapt to data with a low sampling frequency (30 Hz) but also significantly reduce the computational complexity, while maintaining a high recognition accuracy (96.7%).