Committee:

Dr. Maysam Ghovanloo, ECE, Chair , Advisor

Dr. Elliot Moore, ECE, Co-Advisor

Dr. Omer Inan, ECE

Dr. Gregory Abowd, IC

Dr. Fatih Sarioglu, ECE

Dr. Thad Starner, CoC

Abstract: 

Automatic food intake monitoring using wearable sensor-based systems is an alternative to manual self-report methods. Automatic methods aim to quantitatively track aspects related to eating, drinking and/or any form of energy consumption in an effort to encourage healthier dietary behaviors. In this dissertation, a detailed evaluation of research work in the field was undertaken to outline pros and cons of various sensing modalities for on-body use. The most relevant signal processing and machine learning techniques were identified, including best features for acoustic-, image-, and motion-based methods. To address some of the observed research gaps, we focused more on acoustic-based sensing of food intake activities and developed the first real-time swallowing detection algorithm. Following this, we introduced a tracheal activity recognition algorithm based on sub-optimally sampled acoustic signals for energy efficiency purposes. Another observed research gap relates to detecting dietary activities in noisy environments particularly for acoustic-based monitoring systems that are highly affected by background noise. To this effect, we developed a source separation method using semi-supervised non-negative matrix factorization for the enhancement of food intake acoustics in noisy recordings. We also introduced a low-cost template-matching method to detect food intake acoustics in very low signal-to-noise ratio recordings. This research work contributes to the development of a robust, sensor-based, wearable dietary monitoring system. Such a system aims to curtail the growing crisis of obesity, diabetes, eating disorders and other related chronic conditions.