| Popis: |
In the last years, the pervasive use of smartphones and wearable devices, with digital sensors, like accelerometers, gyroscopes, and magnetometers, makes it easy to assess activities of daily living, with several applications and benefits for human beings, as: monitoring elderly patients for the occurrence of falls; monitoring diabetic and Parkinson's disease patients; monitoring of the rehabilitation process of patients, etc. Major of the works published in literature use a single machine lo classify all activities. In this work we propose a separate machine learning process to detect group of activities. The main advantages of our approach are the following enables customizing the sampling window size and overlapping for detecting each activity group, enables a separated adjustment of the parameters and hyperparameters of each classifier, and enables using different signals for detection of each activity group. Therefore, in this work we used different signals, sampling window size, overlapping and classifiers for detecting each activity group. The mean accuracy obtained in activities of daily living detection, using 5-fold-cross validation situated between 97.32% and 99.76%. In literature, using a single classifier, the reported accuracies in detecting activities of daily living situated in the range between 52% to 99.15%. |
