Towards Cross Domain CSI Action Recognition Through One-Shot Bimodal Domain Adaptation

Human action recognition based on WiFi Channel State Information (CSI) has attracted enormous attention in recent years. Although performing well under supervised learning, the recognition model suffers from significant performance degradation when applied in a new domain (e.g. a new environment, a different location, or a new user). To enable the recognition model robust to domains, researchers have proposed various methods, including semi-supervised domain adaptation, unsupervised domain adaptation, and domain generalization. Semi-supervised and unsupervised solutions still require a large number of partially-labeled or unlabeled samples in the new domain, while domain generalization solutions have difficulties in achieving acceptable accuracy. To mitigate these problems, we propose a one-shot bimodal domain adaptation method to achieve cross domain action recognition with much reduced effort. The method contains two key points. One is that it synthesizes virtual samples to augment the training datatset of the target domain, requiring only one sample per action in the target domain. The other is that it regards the amplitude and the phase as two consistent modals and fuses them to enhance the recognition accuracy. Virtual data synthesis is achieved by linear transformation with dynamic domain weights and the synthesis autoencoder. Bimodal fusion is achieved by the fusion autoencoder and feature concatenation under the criterion of consistency. Evaluations on daily activities achieved the average accuracy of 85.03% and 90.53% at target locations, 87.90% and 82.40% in target rooms. Evaluations on hand gestures achieved the average accuracy of 91.67% and 85.53% on target users, 83.04% and 88.01% in target rooms.