Cross-Scenario Device-Free Gesture Recognition Based on Parallel Adversarial Network

Abstract

Wireless sensing has garnered significant attention as a key technique for 6G, as it empowers wireless networks with sensing capabilities. One emerging technology in this domain is device-free gesture recognition (DFGR), which enables the recognition of human gestures by analyzing the influence they exert on the surrounding wireless signals. Deep network based DFGR systems have demonstrated impressive performance thanks to the feature extraction capabilities of deep networks. However, these systems encounter significant performance degradation in cross-scenario conditions, wherein it becomes challenging, and sometimes even impossible, to extract common features that are unrelated to specific working scenarios, particularly when there are substantial differences among the scenarios. To solve this problem, in this paper, we propose and design a parallel adversarial network. Our key idea is to extract common features between the target scenario and each source scenario separately and parallelly, so that we can achieve common features even when the difference between the scenarios is quite large. Specifically, we design adversarial sub-networks for each pair of target and source scenarios to extract their common features and make coarse recognition, develop a similarity evaluation sub-network to estimate the similarity between the target scenario and every source scenario, and fuse the coarse results by leveraging similarity scores to accomplish accurate recognition. We conducted extensive evaluations on two mmWave testbeds and the publicly available Widar3.0 WiFi dataset, and confirmed the effectiveness of the proposed network.