Study and Characterization of the Main Tools for Human Activity Recognition using Accelerometer Sensors

Abstract

Human Activity Recognition (RAH) aims to classify the activities performed by a user collecting data from heterogeneous sensors. The RAH allows the monitoring of user actions, offering services in the area of medical care, in the accompaniment of the elderly, health monitoring, fitness tracking, home and work automation, among others. The RAH can be seen as an Information System composed by three steps: data collection and preprocessing, feature extraction and classification. Despite the abundance of works proposed for this subject, an important issue to be addressed is how to choose the tools and methods to be used in each step of the RAH. This choice is a difficult process, because it involves comparing the results obtained by other works, most of which use private datasets, extract different sets of features, and use different classification algorithms. This paper aims to characterize and compare the main tools, methods and databases for the RAH task. In addition, it aims to provide guidance and guidelines for future research in the area. Experiments were performed in order to identify the main attributes to be used in the classification. It can observed the attributes mean, standard deviation, and variance produce the best models to the classification task.