Cross-domain gesture recognition via WiFi signals with deep learning

Compared with systems rely on wearable sensors, cameras or other devices, WiFi-based gesture recognition systems are convenient, non-contact and privacy-friendly, which have received widespread attention in recent years. In WiFi-based gesture recognition systems, the channel state information (CSI) carried by WiFi signals contains fine-grained information, which is commonly used to extract features of gesture activities. However, since the CSI patterns of the same gesture change across domains, these gesture recognition systems cannot effectively work without retraining in new domains, which will hinder the practical adoption of gesture recognition systems. Therefore, we propose a novel gesture recognition system that can address the issue of cross-domain recognition while achieving higher recognition accuracy for in-domain scenarios. Firstly, we employ CSI ratio and subcarrier selection to effectively eliminate noise from the CSI, and propose a method to reconstruct CSI sequence using low-frequency signals, which can effectively remove irrelevant noise in the high-frequency part and ensure the validity of the data. Next, we calculate the phase difference to explore the intrinsic features of gesture and convert the obtained data into RGB image. Finally, we use Dense Convolutional Network as backbone network, combined with dynamic convolution module, for RGB image recognition. Extensive experiments demonstrate that our proposed system can achieve 99.58% in-domain gesture recognition, and its performance across new person and orientations is 99.15% and 98.31%, respectively.