Algorithm for motion detection and gait classification based on mobile phone accelerometer data

This paper briefly describes the development of information technology tools using biometric data, in particular, human gait parameters. The problems of assessing gait parameters using a mobile phone accelerometer in real conditions are briefly described. The relevance of this research is substantiated in the field of developing algorithms for assessing biometric gait indicators based on data from wearable devices. The main approaches to the processing of wearable device accelerometer data are considered, the main shortcomings and problems in improving the quality of gait parameter estimation are indicated. The algorithm for processing data from a mobile phone accelerometer is described. In the proposed algorithm, the selection of movement patterns during gait in the recorded data is carried out on the basis of statistical information within the “floating” time window (frequency component with the maximum contribution to the spectrum of the accelerometer signal, the duration of the selected time segments), as well as on the basis of the value of the correlation coefficient, selected time segments. At the stage of data segmentation, the time window of searching of movement segments, as well as the allowable thresholds of selecting movements by their duration, change depending on the individual characteristics of the gait and human activity. The classification of the selected segments according to the nature of gait movements is carried out on the basis of a feed-forward neural network. The sigmoid was used as the activation function for the hidden layers, and the normalized exponential function was used for the output layer. The neural network was trained using the gradient backdescent method with cross entropy as an optimization criterion. Due to the selection of segments with a high correlation coefficient, the classification of data shows the quality of distinguishing movements above $95\%$.