In recent years, Body Sensor Networks (BSNs) have been used as the basis of many systems aimed at monitoring bodily parameters ranging from skin temperature, to breathing, to motion. These measurements can then be used to generate additional information related to the monitored subject, such as for heat stress prediction or fall detection. This thesis is concerned with the design, development and realistic evaluation of a BSN-based end-toend posture classification platform using on-body accelerometers. The work is motivated by applications that require stable, sub-second, end-to-end classification of postures, as well as dynamically configurable operation to support exploratory data collection. Classification is performed on-node, thus reducing the amount of data/information transmitted from the wearable nodes to a data sink. The work is experimentally-led, and uses an application case study—on-body monitoring of Explosive Ordnance Disposal (EOD) operatives—to provide context for system requirements and experimentation performed. This thesis provides three main contributions: First, the design of a platform that allows real-time on-body classification of static and dynamic postures—a capability not present in existing work. The specific posture set consists of six static postures (sitting, standing, kneeling, and lying on back, front and one side) and two dynamic postures (walking and crawling), of which kneeling and crawling are not commonly considered in the literature. Classification is performed on a small, light embedded device using a simple easy-to-implement algorithm. The classification algorithm used is a C4.5 decision tree, with a temporal feature (windowed variance) to aid in distinguishing dynamic and static postures. Offline classification accuracy is shown to be 96.3% based on data gathered from subjects in a laboratory environment, and real-time on-node classification accuracy is shown to be comparable to this figure (95.5%). Second, further advance beyond the state of the art is presented through an investigation into posture transitions. Posture transitions cause transient (<1s) posture changes in the classifier output and are shown to reduce classifier accuracy by 2% for every transition / minute for classifiers not specifically designed to handle transitions. Three posture filters that remove such transient posture changes are designed, implemented and tested on experimental data. The best performing filter, Exponentially Weighted Voting (EWV), is shown to reduce posture change events by 75.2% and increase accuracy by 1% (over unfiltered results). Compared to streaming raw data, an event-based posture classification system is shown to reduce transmissions by 98.5% (66-fold reduction). Finally, a broad investigation is presented into the effect of both system-related and training-process factors on the accuracy of machine learning-based posture classifiers. The factors analysed include i) temporal and feature parameters, ii) sensor sampling rate, iii) number of sensors used, iv) posture class aggregation and v) number of subjects used for training. Optimal parameters are determined for the motivating EOD application, with a range of parameter values shown to guide development of other classifiers. Through the novel contributions presented, this thesis provides a solid groundwork for further research in BSN based posture classification systems and simplifies optimisation of machine-learning classifiers for specific posture classification applications.
|Date of Award||2012|
|Supervisor||Elena Gaura (Supervisor) & James Brusey (Supervisor)|