Neural Network Algorithm for Recognizing Human Activity from Radio Signals with Preliminary Motion Detection

This is the authors' abstract. We don't add key points for this paper.

The paper is devoted to the analysis of changes in radio signals caused by human movements, using channel matrices obtained from WiFi devices. The main objective of the study is to develop and study an algorithm capable of classifying various types of activity based on the analysis of radio signals. The proposed algorithm includes a preliminary motion detection stage, which eliminates uninformative data and improves the accuracy of activity recognition. The motion predetection stage includes extracting signal amplitude from channel matrices to analyze motion-induced changes, prefiltering using a simplified Kalman filter to reduce the influence of noise, reducing the dimensionality of data using principal component analysis to simplify processing, and segmenting the signal and estimating statistical parameters (mean and standard deviation) to classify the presence of motion. The classification algorithm involves the use of a convolutional neural network that takes as input multidimensional amplitude matrices concatenated over time and outputs a vector of probabilities for each of the three activity classes. Experimental studies showed that the proposed algorithm with preliminary motion detection achieves an activity classification accuracy of 96.8%, making it applicable to various conditions.