Features Selection for Human Activity Recognition in the Telerehabilitation System

The article researches the problem of human activity recognition in the telerehabilitation system. To recognize the patient's physical activity, the sensors of the smartphone are used: accelerometer and gyroscope. The term of telerehabilitation and an exemplary set of exercises as part of a rehabilitation event are considered. Such tasks as determining the permissible value of frequency and classification accuracy and also features select to reduce computational complexity are solved. The classification frequency value is proposed to take into account the type of activity and the patient's health group for evaluating the correctness of independent implementation of rehabilitation events by the patient. An algorithm for selecting a subset of informative features is described. An experiment is carried out to select a subset of informative features is necessary for the classification of physical activity in the telerehabilitation system, taking into account the influence of features on the classification accuracy and computational complexity in their calculation. Comparison of classification results using a feature vector and using a subset of informative features is performed.