LAGER: Label-Free Domain-Adaptive Wireless Gesture Recognition via Latent Feature Alignment and Augmentation

As a non-verbal form of communication, gestures convey information through bodily movements and postures. Gesture recognition provides a more intuitive and natural human-computer interaction (HCI) experience, making it an integral component of the field of HCI. Recently, Wi-Fi-based gesture recognition has become a popular direction in research and applications due to its low cost, privacy-friendly nature, and convenience. However, due to the differences in the distribution of gesture data between the known and target environments, deploying the gesture recognition model in a new environment may induce high costs in annotating data labels and retraining the model. To achieve a cost-effective transferable gesture recognition model, we propose an efficient Wi-Fi-based gesture recognition domain-adaptive method (LAGER) that can maintain high recognition accuracy in a new environment without the need for labeled samples. Specifically, LAGER divides the cross-domain Wi-Fi gesture recognition problem into two inter-related sub-problems, where we iteratively apply a pseudo-label guided feature alignment and feature augmentation method in a latent space by leveraging the wisdom of unsupervised domain adaptation. To minimize the negative impact of erroneous pseudo-labels in the early training stage, we introduce a pre-heated training technique that separates the training process into two parts associated with different training strategies. We evaluate our method on the Widar3.0 dataset and compare the performance under various cross-domain settings with several representative benchmark methods. The proposed LAGER evidently outperforms all the benchmark methods. In particular, compared to the cross-domain recognition method used in Widar3.0, the proposed LAGER achieves 6.89% to 7.83% higher average accuracy in different cross-domain experiments considered.