Hierarchical classifier approach to physical activity recognition via wearable smartphone tri-axial accelerometer.

Physical activity recognition has emerged as an active area of research which has drawn increasing interest from researchers in a variety of fields. It can support many different applications such as safety surveillance, fraud detection, and clinical management. Accelerometers have emerged as the most useful and extensive tool to capture and assess human physical activities in a continuous, unobtrusive and reliable manner. The need for objective physical activity data arises strongly in health related research. With the shift to a sedentary lifestyle, where work and leisure tend to be less physically demanding, research on the health effects of low physical activity has become a necessity. The increased availability of small, inexpensive components has led to the development of mobile devices such as smartphones, providing platforms for new opportunities in healthcare applications. In this study 3 subjects performed directed activity routines wearing a smartphone with a built in tri-axial accelerometer, attached on a belt around the waist. The data was collected to classify 11 basic physical activities such as sitting, lying, standing, walking, and the transitions in between them. A hierarchical classifier approach was utilised with Artificial Neural Networks integrated in a rule-based system, to classify the activities. Based on our evaluation, recognition accuracy of over 89.6% between subjects and over 91.5% within subject was achieved. These results show that activities such as these can be recognised with a high accuracy rate; hence the approach is promising for use in future work.