Human Activity Recognition System Using Artificial Neural Networks

Population aging and the increasing costs of health care, especially for the elderly affected by chronic diseases, requires new medical assistance strategies that makes it possible to monitor these people remotely and provide reliable information on their routines. In this context, human activity recognition (HAR) systems are an important element to overcoming the problem. Therefore, this paper proposes a HAR system prototype containing a multilayer perceptron (MLP) as a classifier. The model hyperparameters were selected using a publicly available dataset. Then, data was collected from accelerometers and gyroscopes embedded in wearable devices of 15 subjects while performing six basic activities (walking, sitting, lying down, standing, walking upstairs and walking downstairs). The system reached an average accuracy of 90.74% and weighted F-measure of 90.03% based on leave-one-subject-out cross-validation.