Enhancing Human Activity Recognition Performance in Small-Sample Wi-Fi Sensing Using Data Augmentation Methods

Activity recognition based on Wi-Fi channel state information(CSI) has exhibited promising outcomes. However, numerous prevailing CSI-based activity recognition systems heavily lean on substantial datasets for model training. In reality, procuring CSI data about human behavior demands significant resources. Consequently, the challenge emerges in enhancing Wi-Fi perception under circumstances of low sampling rates. This research endeavors to investigate avenues for augmenting the available dataset. Our findings substantiate that the predicament of limited samples in Wi-Fi environments can be effectively alleviated by extending CSI from a one-dimensional time series to a two-dimensional recurrence plot (RP), coupled with the integration of data enhancement methodologies commonly deployed in computer vision. The results highlight that combining RP transformations with the horizontal flip technique achieves an 86% recognition rate using only 5% of the training samples, and a recognition rate of approximately 96% using a mere 10% of the training samples. Moreover, an outstanding recognition accuracy of 99.5% is attained with only half of the training samples, surpassing the capabilities of existing solutions in this field. We will publish our code at https://github.com/shengjunzi/Enhancing-Human-Activity.