Cross-Scenario Device-Free Activity Recognition Based on Deep Adversarial Networks

Device-free activity recognition (DFAR) is an emerging technique which could infer human activities by analyzing his/her influence on surrounding wireless signals. It may empower wireless networks with the additional sensing ability. Existing studies have achieved reasonable accuracy in a pre-trained scenario. However, due to the feature shift incurred by different radio environments, a system typically achieves poor performance in a new scenario. Generally, retraining a system is laborious or even impossible in practical applications, since we have very few number of or even no labeled training samples in a new scenario. Therefore, how to realize cross-scenario DFAR in an unsupervised manner becomes an urgent problem to solve. To address this challenge, in this paper, we develop a deep learning network to guide the sample features of the target scenario shift to those of the source scenario without using any label information of the target scenario. Specifically, we develop a maximum-minimum adversarial approach to move the target features to the distribution of the source features, and design a center alignment strategy to further shift the target features to the distribution center. Benefit from the shifted features, extensive experimental results on a mmWave testbed demonstrate the effectiveness of the developed framework.