Cross-domain human activity recognition using reconstructed Wi-Fi signal

Abstract

Recent studies have shown that Wi-Fi channel state information (CSI) based approaches for human activity recognition (HAR) is successful. However, the performance of these approaches often deteriorate when deployed to a new industrial environment. To solve this problem without retraining, we present a novel Wi-Fi CSI tensor based cross-domain HAR approach (TensFi). Specifically, activity-related CSI is first separated from the original CSI through an ensemble empirical mode decomposition (EEMD) algorithm. Then, the sparse signal representation (SSP) algorithm is used to extract partial CSI sub-carriers that are more relevant to human activities. Furthermore, the sparse CSI is modeled as phase-integrated CSI (PI-CSI) and further reconstructed as a CSI tensor with unique decomposition. After that, CANDECOMP/PARAFAC (CP) is used to decompose the reconstructed CSI tensor. Finally, a multi-head self-attention based gated temporal convolutional network (MAGTCN) is designed to capture the features of decomposed CSI tensor and then make the final activity decision. Experimental results show that TensFi can achieve good cross-domain generalization performance.