Abstract:
Recently, Automated Human Activity Recognition has been extensively used for longterm health monitoring of healthy individuals and to provide assisted living for the
elderly. Long-term health monitoring systems have been successfully implemented for
the prevention of chronic diseases like heart disease, obesity, workers syndrome, and
other diseases related to sedentary lifestyle.
Activities of Daily Living such as sitting, standing, walking, working in office, jogging,
and running etc. can be efficiently classified using sensors such as 3D accelerometers,
gyroscopes, and magnetometers. Such systems have demonstrated very high
classification accuracies for activities performed for longer durations of time. However,
these systems are unable to detect and classify transitory activities where the subject
switches from one basic activity to another. For example, if a subject stands up from
the chair to walk out of the room and then goes downstairs to reach ground floor of the
office building has transitions from stationary sitting to standing and then walking,
walking to going downstairs and then walking again and these transitions may not be
classified correctly by existing automated human activity recognition systems since the
models are trained using nonrealtime segmented data for each individual activity class.
This research aims to develop a system for the detection of transitory activities. A
Mobile phone-based accelerometer is used to record these activities from the chest of
subject through MyNeuroHealth application. Data is collected, pre-processed, and
classified into different activity classes. This data is used to train Artificial Neural
Network to classify transitory activities. The proposed system achieved an accuracy of
more than 50% with real-time data. Furthermore, it is also observed that using two
accelerometers for collecting the movement data can enhance classification accuracy
to 65%. Given that little or no work has been done in this dimension of HAR, this
research may be extended to improve the accuracy of HAR for real-time automated
long term health monitoring systems.
