AIGC for RF-Based Human Activity Sensing

Radio frequency (RF) sensing has been considered as an effective approach to human perception of nonintrusive and high-privacy scenarios. However, the existing wireless sensing techniques mostly rely on extensive labeled RF sensing data for offline training, while wireless sensory data collection is highly time consuming and costly. To ridge this gap, we investigate the problem of generalized dataset augmentation with an artificial intelligence (AI) generated content (AIGC) approach, termed RF-AIGC, for wireless sensing, which can not only purposefully generate new RF sensing data but reduce the data collection cost by augmenting a limited training dataset with synthesized RF data. We propose a conditional recurrent generative adversarial network (termed RF-CRGAN) to generate labeled synthetic RF data for specified human activities for multiple wireless sensing platforms, such as WiFi, radio-frequency identification (RFID), and millimeter wave (mmWave) radar. We also propose a holistic quantitative method to help evaluate and explain the effects of the synthesized data. The experimental results demonstrate that the proposed approach can effectively enhance the diversity of training data and achieve similar performance as real data.