ImgFi: A High Accuracy and Lightweight Human Activity Recognition Framework Using CSI Image

Although human activity recognition (HAR) based on WiFi channel state information (CSI) has been widely studied, as the CSI signal is susceptible to the external environment, a recognition framework with high accuracy and low complexity remains a key and difficult problem. To improve WiFi sensing performance and reduce model complexity, ImgFi, a gesture recognition system, is proposed, which includes a conversion module and a recognition module. The conversion module is used to convert the CSI to images, and the key difference of the proposed ImgFi is using signal variations in arbitrarily different timestamps as feature extraction units rather than single signal points. This conversion can also take advantage of the powerful ability of the convolutional neural network (CNN) in the field of image recognition. Five CSI imaging approaches are introduced in ImgFi to realize the CSI conversion while a CNN is designed to recognize the activity from the converted images. To verify the proposed framework, we make some tests in our laboratory and get a dataset. The tests on four datasets under different contexts indicate that ImgFi with only three convolutional layers achieves 99.5% recognition accuracy. That is, our method has higher accuracy and lower complexity, which is important to related applications, such as smart homes and virtual reality. We will publish our code at https://github.com/shengjunzi/ImgFi .