Performance Assessment of Deep Neural Network on Activity Recognition in WiFi Sensing

WiFi sensing technology has received growing attention because of its widely available infrastructure and capability to support Internet of Things (IoT) driven smart-world applications. With the advance of machine learning and computing techniques, deep learning techniques have been applied to WiFi sensing, supporting a variety of applications. In this paper, we address the issue of understanding the ability of deep learning to handle complex and diverse sensing data collected from different WiFi sensing environments. To this end, we propose a generic framework that considers the sensing activity, sensing target, sensing environment, and related configuration. Based on the framework, we design our deep neural network (DNN)-based approach to carry out activity recognition based on WiFi sensing data. We select seven datasets to represent diverse WiFi sensing applications. Via extensive performance evaluation, the results confirm that our proposed DNN-based approach can achieve an average accuracy of 93.66 % in different settings with respect to diverse target, activity, environment, and configuration.