WiSDA: Subdomain Adaptation Human Activity Recognition Method Using Wi-Fi Signals

Abstract

Human activity recognition based on Wi-Fi signals has become one part of integrated sensing and communications, which has promising application prospects. Detecting activities across different domains is an important and challenging problem. To reduce model complexity and improve recognition accuracy, we propose a novel approach to realize activity recognition across domains, named WiSDA. The proposed WiSDA contains two parts: data augmentation and a deep learning model. The recursive plots method is employed as the data augmentation to transform Wi-Fi channel state information into images, which can take advantage of the image recognition ability of the latter deep learning model. The proposed learning model utilizes weighted cosine similarity to align feature distributions among sub-domains activated by a deep network layer across different domains, thereby a domain-independent feature representation is generated. Based on this representation, WiSDA can make the recognition decision independent of domains, then the cross-domain recognition accuracy is increased. The numerical results illustrate that WiSDA achieves higher recognition accuracy and has lower complexity. The cross-domain recognition accuracy ranges from 89% to 93% with offline pre-training. Enhancing the pre-trained WiSDA with limited samples boosts cross-domain recognition accuracy to 97%.