Device-Free Human Activity Recognition by Radio Engineering Systems Based on Neural Network Technologies

Currently, the development of information technology places ever higher demands on the accuracy and speed of solving various applied problems. The paper considers the problem of recognizing human activity without using body sensors using channel matrices of wireless devices. Recognizing human activity is an important practical task in the development of human-machine interaction systems. In recent years, the number of studies on recognizing human activity has increased significantly due to the increased availability of various sensors and a decrease in their cost. That is why it is proposed to use information on how the signal propagates from the transmitter to the receiver - channel matrices. The study of channel matrices is a relevant area of scientific research to improve the accuracy and speed of work when solving problems of positioning, recognition, etc. The paper proposes the use of neural network technologies in view of the fact that an explicit analytical relationship between activities and multidimensional channel matrices is unknown. Thus, four types of neural network models have been developed: LSTM, 1D-CNN, 1D-LSTM, BLSTM. To train the models, a training sample was formed, including information on the activities of five subjects. The experimental studies also used a test sample, including channel matrices collected with each other, in a similar room, also with the participation of five subjects. Four neural network models are proposed, the accuracy of which reaches 98.56% at the training stage. The most accurate model, according to the results of the studies, was the BLSTM model with an accuracy value of 86.26%. Moreover, the accuracy of all the tests is not much, but higher than its competitors. BLSTM more accurately recognizes the fact that a person is sitting or lying. A standing person is more accurately recognized by the 1D-LSTM model. But walking and running in place are better recognized by the LSTM and 1D-CNN models, respectively. In conclusion, we can conclude that the use of channel matrices to solve the problem of recognizing human activity allows you to achieve highly accurate results while maintaining the ability to scale.