Poster: RF-HGR: Domain-independent Few-Shot Recognition for Unseen Gestures with RFID
Haofan Cai, Chen Qian · 2022
RFID-based Human Gesture Recognition (HGR) has gained much attention and become a promising solution for device-free human-computer interaction (HCI) in recent years. However, existing RFID sensing suffers from limited scalability as the system needs to be re-trained whenever unseen gestures are introduced, which causes overheads of data collection and re-training. Meanwhile, cross-domain sensing may also fail as the correlation between gestures and induced variations on wireless signals will change when a different domain (i.e., environment or user) is involved. In this paper, we propose a RFID-based HGR system named RF-HGR, which can recognize unseen classes in new domains with only a few labeled samples. Specifically, RF-HGR employs a lightweight few-shot learning (FSL) framework based on fine-tuning domain adaptation to eliminate model re-training overhead. We establish a real-world prototype using commercial off-the-shelf (COTS) RFID devices and the preliminary evaluation results show that RF-HGR with three-, five-, seven-shot learning can recognize novel classes in unseen domains with an accuracy of 50.6%, 65.7% and 72.8% respectively.