Date of Award

8-2022

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering (Holcomb Dept. of)

Committee Chair/Advisor

Adam Hoover

Committee Member

Harlan Russell

Committee Member

Jacob Sorber

Committee Member

Ian Walker

Abstract

This dissertation describes a “smart dining table” that can detect and measure consumption events. This work is motivated by the growing problem of obesity, which is a global problem and an epidemic in the United States and Europe. Chapter 1 gives a background on the economic burden of obesity and its comorbidities. For the assessment of obesity, we briefly describe the classic dietary assessment tools and discuss their drawback and the necessity of using more objective, accurate, low-cost, and in-situ automatic dietary assessment tools. We explain in short various technologies used for automatic dietary assessment such as acoustic-, motion-, or image-based systems. This is followed by a literature review of prior works related to the detection of weights and locations of objects sitting on a table surface. Finally, we state the novelty of this work.

In chapter 2, we describe the construction of a table that uses an embedded grid of load cells to sense the weights and positions of objects. The main challenge is aligning the tops of adjacent load cells to within a few micrometer tolerance, which we accomplish using a novel inversion process during construction. Experimental tests found that object weights distributed across 4 to 16 load cells could be measured with 99.97±0.1% accuracy. Testing the surface for flatness at 58 points showed that we achieved approximately 4.2±0.5 um deviation among adjacent 2x2 grid of tiles. Through empirical measurements we determined that the table has a 40.2 signal-to-noise ratio when detecting the smallest expected intake amount (0.5 g) from a normal meal (approximate total weight is 560 g), indicating that a tiny amount of intake can be detected well above the noise level of the sensors.

In chapter 3, we describe a pilot experiment that tests the capability of the table to monitor eating. Eleven human subjects were video recorded for ground truth while eating a meal on the table using a plate, bowl, and cup. To detect consumption events, we describe an algorithm that analyzes the grid of weight measurements in the format of an image. The algorithm segments the image into multiple objects, tracks them over time, and uses a set of rules to detect and measure individual bites of food and drinks of liquid. On average, each meal consisted of 62 consumption events. Event detection accuracy was very high, with an F1-score per subject of 0.91 to 1.0, and an F1 score per container of 0.97 for the plate and bowl, and 0.99 for the cup. The experiment demonstrates that our device is capable of detecting and measuring individual consumption events during a meal.

Chapter 4 compares the capability of our new tool to monitor eating against previous works that have also monitored table surfaces. We completed a literature search and identified the three state-of-the-art methods to be used for comparison. The main limitation of all previous methods is that they used only one load cell for monitoring, so only the total surface weight can be analyzed. To simulate their operations, the weights of our grid of load cells were summed up to use the 2D data as 1D. Data were prepared according to the requirements of each method. Four metrics were used to evaluate the comparison: precision, recall, accuracy, and F1-score. Our method scored the highest in recall, accuracy, and F1-score; compared to all other methods, our method scored 13-21% higher for recall, 8-28% higher for accuracy, and 10-18% higher for F1-score. For precision, our method scored 97% that is just 1% lower than the highest precision, which was 98%.

In summary, this dissertation describes novel hardware, a pilot experiment, and a comparison against current state-of-the-art tools. We also believe our methods could be used to build a similar surface for other applications besides monitoring consumption.

Author ORCID Identifier

0000-0002-7103-7164

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.