Human Activity Recognition System For Long Term

Abstract

The study of Human Activity Recognition (HAR) for Long Term Health Monitoring (LTHM) has gained significant importance for its wide range of applications. These applications range from sports and rehabilitation sciences to assisted living for older people. In addition to that LTHM is an efficient solution for the prevention of lifestyle diseases like stroke, heart failure, and various health problems that occur due to prolonged inactivity. With the increased availability of accelerometer sensors embedded in mobile phones, we can efficiently explore the Activities of Daily Living (ADLs) of an individual. This research aims to develop a LTHM system for evaluating ADLs of a person using a mobile phone-based accelerometer sensor and the ‘MyNeuroHealth’ application. Data collected in an unconstrained environment by various individuals throughout the day to create templates of ADLs. Collected data is prepared and preprocessed by assigning hourly labels to the ADLs, encoding categorical values and random sampling of data. This data is used for training the machine learning model and for classifying activities according to their energy expenditure or user exhaustion levels. Collected dataset further extended to daily, weekly, and monthly basis to provide long-term health profiling (LTHP). 23 types of basic, complex and transitional activities were evaluated for each day. The results show that an Artificial Neural Network (ANN) can efficiently identify and detect ADLs with more than 90% accuracy. Person independent ADLs templates for weeks 1, 2, 3 and 4 achieved an accuracy of 89, 96, 93 and 89 percent correspondingly. On the other hand, person dependent ADLs templates from various walks of life achieved on average 94% accuracy. Person independent ADLs templates for weeks 1, 2, 3 and 4 achieved an accuracy of 89%, 96%, 93% and 89% percent correspondingly.