Date of Award

May 2020

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical and Computer Engineering (Holcomb Dept. of)

Committee Member

Adam W. Hoover

Committee Member

Richard R. Brooks

Committee Member

Richard E. Groff

Abstract

This research considers training a deep learning neural network for segmenting and classifying eating related gestures from recordings of subjects eating unscripted meals in a cafeteria environment. It is inspired by the recent trend of success in deep learning for solving a wide variety of machine related tasks such as image annotation, classification and segmentation. Image segmentation is a particularly important inspiration, and this work proposes a novel deep learning classifier for segmenting time-series data based on the work done in [25] and [30]. While deep learning has established itself as the state-of-the-art approach in image segmentation, particularly in works such as [2],[25] and [31], very little work has been done for segmenting time-series data using deep learning models.

Wrist mounted IMU sensors such as accelerometers and gyroscopes can record activity from a subject in a free-living environment, while being encapsulated in a watch-like device and thus being inconspicuous. Such a device can be used to monitor eating related activities as well, and is thought to be useful for monitoring energy intake for healthy individuals as well as those afflicted with conditions such as being overweight or obese.

The data set that is used for this research study is known as the Clemson Cafeteria Dataset, available publicly at [14]. It contains data for 276 people eating a meal at the Harcombe Dining Hall at Clemson University, which is a large cafeteria environment. The data includes wrist motion measurements (accelerometer x, y, z; gyroscope yaw, pitch, roll) recorded when the subjects each ate an unscripted meal. Each meal consisted of 1-4 courses, of which 488 were used as part of this research. The ground truth labelings of gestures were created by a set of 18 trained human raters, and consist of labels such as ’bite’ used to indicate when the subject starts to put food in their mouth, and later moves the hand away for more ’bites’ or other activities. Other labels include ’drink’ for liquid intake, ’rest’ for stationary hands and ’utensiling’ for actions such as cutting the food into bite size pieces, stirring a liquid or dipping food in sauce among other things. All other activities are labeled as ’other’ by the human raters. Previous work in our group focused on recognizing these gesture types from manually segmented data using hidden Markov models [24],[27]. This thesis builds on that work, by considering a deep learning classifier for automatically segmenting and recognizing gestures.

The neural network classifier proposed as part of this research performs satisfactorily well at recognizing intake gestures, with 79.6% of ’bite’ and 80.7% of ’drink’ gestures being recognized correctly on average per meal. Overall 77.7% of all gestures were recognized correctly on average per meal, indicating that a deep learning classifier can successfully be used to simultaneously segment

and identify eating gestures from wrist motion measured through IMU sensors.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.