Abstract/Summary

Using multiple Inertial Measurement Units (IMU) in movement analysis will not only be useful in increasing the classification accuracy of the movement data, but also reduce the computational complexity in classification algorithms as there is no need to process an increased number of features generated from a single sensor. However, wearing sensor devices every day on the same place with the same orientation is a key requirement for the data analysis purpose. To facilitate this, sensor devices can be mounted into clothing, as it is an ideal platform to cater these miniature devices. There are research studies conducted with sensors mounted into clothing such as smart garments, tight-fitting clothing and loose clothing (everyday wear clothing). Data validations are available between tight-fitting clothing-mounted sensor data and body-mounted sensor data, focusing mainly on limited set of activities or sensors. The main focus of this research was to investigate the possibility of using loose clothing-mounted sensors in monitoring human movement patterns in a home based healthcare monitoring system, while validating how the loose clothing-mounted sensor data correlate with body-mounted sensor data with respect to different activities. In order to quantify and understand human movements in this research, time synchronised wearable sensors were mounted into loose clothing. This whole research was based on three datasets and they were used to conduct four sub analyses based on different types of human movement patterns to achieve the main goal. First analysis was based on data collected from Actigraph sensors from both body and clothing and the sensors were near waist, thigh and ankle/ lower-shank. This study validated the data between clothing and body mounted sensor data across various static and dynamic activities with respect to each sensor pairs. These validations were based on correlation coefficient values with respect to the accelerometer data pairs for different activities i.e. ‘standing’, ‘sitting’, ‘sitting on a bus’, ‘walking’ and ‘running’. Promising correlations were observed (especially with static activities) with this dataset and the second dataset was collected from body and clothingmounted lightweight IMU sensors. These data were analysed based on correlation coefficient values with respect to the inclination angle changes over ‘gait’ cycles. In addition to the correlation coefficient values, the data were analysed using different types of plots such as phase portraits and 3D plots. From these plots, it was noted that important features such as Mid-Stance (MS), Initial Contact (IC) and Toe Off (TO) points can be recognised by clothing data and they can be used to analyse ‘walking’ data in detail. Moreover, the third semi-natural dataset was collected from clothing-mounted sensors to check whether they can be used to implement posture and activity classifiers. These classifiers that were based on both Machine Learning (ML) and Deep Learning (DL) approaches with relevant selected features, also showed reasonably high classification accuracies. By taking into account all these promising observations, this thesis can be concluded that loose clothing-mounted sensor data can be used productively in movement analysis.