Human Activity Recognition Based on CSI fragment with Action-value Method

Abstract

The application of Human Activity Recognition (HAR) technology makes human life more convenient. As an emerging HAR technology, WiFi based wireless sensor can sense the state of the target, such as body movement, gesture, position and so on. Aiming at the problem that the current WIFi based HAR methods need more training samples and have low real-time performance, this paper proposes a novel HAR method based on action-value method. In this method, each complete Channel State Information (CSI) sample signal of each activity is sliced into piece samples, and these piece samples are trained and tested to improve the real-time performance and reduce the number of training samples. The piece sample and the sequence of piece samples are respectively regarded as the state information and environment, and the classification of each piece sample is regarded as the execution of the classification-action. A deep neural network is used to simulate the reward of classification-actions in each state, and the recognition model is established by the action-value method. We tested our approach on SignFi data set. The highest recognition accuracy rate of active is 99% and 91% respectively.