Human Activity Recognition based on WiFi Signal Using Deep Neural Network

Human activity recognition can be applied to various applications, ranging from intelligent security to medical care. Action recognition via WiFi had caught great attention due to its ubiquity, non-occlusion and privacy preservation. Movement of human body can cause fluctuations in the wireless signal reflection, resulting in fluctuation of channel state information (CSI), which is computed from the received WiFi signal containing detailed information of current channel conditions for different sub-carriers and spatial streams, and this can be used to classify human activities. In this paper, we focus on the human activity recognition problem based on the features extracted from CSI data using deep neural network. First, the application of human activity classification as well as data preprocessing techniques are introduced in details, and then a fully connected network is developed as the baseline network model. Furthermore, the weight sharing technique is introduced with the similarity network structure, and a new hybrid loss containing the classification loss and similarity loss is proposed. Experiment results are also included and show that the proposed network can achieve higher classification accuracy than the baseline model.