Date of Award

5-2014

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Electrical Engineering

Advisor

Hoover, Adam W

Committee Member

Brooks, Richard R

Abstract

Advances in body sensing and mobile health technology have created new opportunities for empowering people to take a more active role in managing their health. Measurements of dietary intake are commonly used for the study and treatment of obesity. However, the most widely used tools rely upon self-report and require considerable manual effort, leading to underreporting of consumption, non-compliance, and discontinued use over the long term. We are investigating the use of wrist-worn accelerometers and gyroscopes to automatically recognize eating gestures. In order to improve recognition accuracy, we studied the sequential ependency of actions during eating. In chapter 2 we first undertook the task of finding a set of wrist motion gestures which were small and descriptive enough to model the actions performed by an eater during consumption of a meal. We found a set of four actions: rest, utensiling, bite, and drink; any alternative gestures is referred as the other gesture. The stability of the definitions for gestures was evaluated using an inter-rater reliability test. Later, in chapter 3, 25 meals were hand labeled and used to study the existence of sequential dependence of the gestures. To study this, three types of classifiers were built: 1) a K-nearest neighbor classifier which uses no sequential context, 2) a hidden Markov model (HMM) which captures the sequential context of sub-gesture motions, and 3) HMMs that model inter-gesture sequential dependencies. We built first-order to sixth-order HMMs to evaluate the usefulness of increasing amounts of sequential dependence to aid recognition. The first two were our baseline algorithms. We found that the adding knowledge of the sequential dependence of gestures achieved an accuracy of 96.5%, which is an improvement of 20.7% and 12.2% over the KNN and sub-gesture HMM. Lastly, in chapter 4, we automatically segmented a continuous wrist motion signal and assessed its classification performance for each of the three classifiers. Again, the knowledge of sequential dependence enhances the recognition of gestures in unsegmented data, achieving 90% accuracy and improving 30.1% and 18.9% over the KNN and the sub-gesture HMM.

Share

COinS