Deep learning in WiFi CSI-based human activity recognition

Abstract

As one of the most common signals in people’s daily life, WiFi signal is widely used in human activity recognition tasks in recent years. Unlike visionbased human action recognition methods, WiFi-based methods are able to recognize occluded human actions due to the penetration and reflection nature of WiFi signal. Despite of the advantage of WiFi signal, there is still a lack of public datasets which consider occlusion in human action comprehensively. Hence, we construct a WiFi-based CSI human activity recognition dataset with commodity WiFi devices. The dataset contains ten classes of actions and three different occlusion scenarios. Based on the proposed dataset, we evaluate the accuracy and robustness of the state-of-the-art WiFi-based deep learning models. Furthermore, we examine the impact of occlusion on WiFi-based human activity recognition and find that the occlusion is a significant factor in improving the diversity of the dataset.