Wearable sensor systems composed of small and light sensing nodes have the potential to revolutionise healthcare. While uptake has increased over time in a variety of application areas, it has been slowed by problems such as lack of infrastructure and the functional capabilities of the systems themselves. An important application of wearable sensors is the detection of falls, particularly for elderly or otherwise vulnerable people. However, existing solutions do not provide the detection accuracy required for the technology to gain the trust of medical professionals. This thesis aims to improve the state of the art in automated human fall detection algorithms through the use of a machine learning based algorithm combined with novel data annotation and feature extraction methods. Most wearable fall detection algorithms are based on thresholds set by observational analysis for various fall types. However, such algorithms do not generalise well for unseen datasets. This has thus led to many fall detection systems with claims of high performance but with high rates of False Positive and False Negative when evaluated on unseen datasets. A more appropriate approach, as proposed in this thesis, is a machine learning based algorithm for fall detection. The work in this thesis uses a C4.5 Decision Tree algorithm and computes input features based on three fall stages: pre-impact, impact and post-impact. By computing features based on these three fall stages, the fall detection algorithm can learn patterns unique to falls. In total, thirteen features were selected across the three fall stages out of an original set of twenty-eight features. Further to the identification of fall stages and selection of appropriate features, an annotation technique named micro-annotation is proposed that resolves annotation-related ambiguities in the evaluation of fall detection algorithms. Further analysis on factors that can impact the performance of a machine learning based algorithm were investigated. The analysis defines a design space which serves as a guideline for a machine learning based fall detection algorithm. The factors investigated include sampling frequency, the number of subjects used for training, and sensor location. The optimal values were found to be10Hz, 10 training subjects, and a single sensor mounted on the chest. Protocols for falls and Activities of Daily Living (ADL) were designed such that the developed algorithms are able to cope under a variety of real world activities and events. A total of 50 subjects were recruited to participate in the data gathering exercise. Four common types of falls in the sagittal and coronal planes were simulated by the volunteers; and falls in the sagittal plane were additionally induced by applying a lateral force to blindfolded volunteers. The algorithm was evaluated based on leave one subject out cross validation in order to determine its ability to generalise to unseen subjects. The current state of the art in the literature shows fall detectors with an F-measure below 90%. The commercial Tynetec fall detector provided an F-measure of only 50% when evaluated here. Overall, the fall detection algorithm using the proposed micro-annotation technique and fall stage features provides an F-measure of 93% at 10Hz, exceeding the performance provided by the current state of the art.
|Date of Award||2013|
|Supervisor||James Brusey (Supervisor) & Elena Gaura (Supervisor)|
- fall detection
- remote patient monitoring
- wearable sensors