Classification of human activities by smart device measurements

Abstract

The prevalence of activity detectors in users’ personal mobile devices has been incorporated into an increasing interest in research into physical function recognition (HAR - Human Activity Recognition). With this research interest, different enterprises developed HAR systems working with measurement devices and still work on this subject. Although many HAR systems have been developed, there are still concrete practical limits. This situation is improved with modern techniques such as machine learning. A properly trained machine learning model predicts human activity from measured data. The data was measured at certain time intervals by sensors on smartphones. These different machine learning architectures were trained on sensor data that detected human activities, and their accuracy was calculated. A HAR system that predicts human activity is constructed separately with five approaches. KNN, Random Forest, Decision Tree, MLP and Gaussian Naive Bayes algorithms were used, and KNN produced the most accurate results.