Abstract/Summary

Physical inactivity is a leading risk factor in public health and inactive people are more vulnerable to having non-communicable diseases (NCDs), for example, autoimmune diseases, strokes, most heart diseases, diabetes, chronic kidney disease, and others. In addition, levels of physical activity may be an indicator of health problems in older adult individuals, a particular problem in many societies where there is a growing ratio of old adults age 65 and over. Identifying levels of physical activity may have a significant effect on fitness and reducing healthcare costs in the future. Thus, finding approaches for measuring the individuals’ activities is an important need, in order to provide information about their quality of life and to examine their current health status. This thesis explores the possibility of using low-cost wearable accelerometer based inertial sensors to determine activities during daily living. Two data sources were used for this investigation. The first was a locally collected data set recorded from individuals with Parkinson’s disease in their own homes where they were asked to stand up from their favourite chair and then do different daily activities (Bridge data set). The second was a data set collected in a movement laboratory of the Fredrich-Alexander university and measures 19 participants doing daily activities (sit, stand, washing dishes, sweeping, walking, etc) in controlled conditions (Benchmark data set). Both studies used accelerometer based measurements as these are widely used in wearable and portable technologies such as smartphones, and are now finding use in health care applications. Two areas of research are considered. In the first, accelerometer data were considered in relation to the surface of a sphere of radius 1g (i.e. magnitude of the acceleration due to earth gravitate). This research looked at sensor placement, window size and novel features based on the ‘gravity sphere’. Decision Trees and Na¨ıve Bayes classifiers were used as a baseline classifier on both data sets and k-Nearest Neighbour was used on the Bench Mark data set only. The classification results of a small set of activities of a single individual from first data set show that Na¨ıve Bayes (NB) had a better overall accuracy rate than Decision Trees (DTs), where the results are 85.41% and 78.56% for both NB and DTs respectively. The second area of work considered the possibility of using models of the dynamic system of the human movement as the basis for movement classification. Data from the accelerometers were used to evaluate two approaches that exploited the modelling capacity of a system identification algorithm. The two methods, which are called Prediction Measuring (PM) and Model Matching (MM), used the recursive least square method to identify a model for each class (activity). The Benchmark data set was used to verify the proposed methods. PM method achieved better classification accuracy comparing to MM method, with 71% and 59% respectively.