Project Summary/Abstract
Overeating and unhealthy eating are often associated with various health risk conditions such as obesity, high
blood pressure, and some chronic diseases. To get a better understanding of overeating and unhealthy eating,
researchers often rely on self-reports provided by individuals. Suggestions regarding changing lifestyle is often
provided based on observations from these self-reports. However, it is well known that self-reports can be
erroneous and subject to reporting biases. Thus, an objective way to measure the eating activity and validating
self-reports is necessary. Recently, there has been growing interest in moving beyond self-reports and
monitoring the eating activity automatically. To monitor automatically, and in real time, researchers have looked
at using sensor data from wrist worn devices, neck-worn devices, or ear-worn devices to automatically detect
eating. These devices often enable capturing the eating periods. However, these devices seldom capture
images, thus limiting the possibility of visually confirming the consumed food and their quantity.
With the increasing popularity of wearable cameras, it is gradually becoming possible to capture the eating
activities and associated context automatically and without any user intervention. Advances in machine learning
enables automatically extracting eating related information from these captured images. However, wearable
cameras often capture more information than necessary, like capturing bystanders. This unnecessary
information capturing reduces participant's willingness to wearing the camera. Currently, no camera exists that
can capture the eating activity and at the same time limit capturing unnecessary information. Obfuscating the
unnecessary information might increase participant's willingness to wear the camera. However, it is unclear if
and which obfuscation technique will increase participant's willingness to don the wearable camera and at the
same time ensure automatic context determination. In this project, we will determine the possibility of using
machine learning to detect eating in videos and identify the obfuscation technique that can allow detecting the
eating activity without collecting unnecessary information.
To this end, first we will develop an activity detection algorithm that will allow detecting the eating activity using
data from an IR sensor array and RGB images. Next, we will test various obfuscation methods in a cross-over
trial and select the best obfuscation method based on the greatest participant acceptability. We will then deploy
the eating detection algorithm with the best obfuscation approach on a novel wearable camera that has an
infrared sensor array. We will use this camera to test the possibility of detecting eating in a real-world setting. To
validate our algorithm, we will ask people to confirm or refute predicted eating and non-eating moments. We will
compare the performance of this algorithm against both real-time user response and 24-hour dietary recall to
objectively evaluate the algorithm's performance. Our proposed system will improve current research practices
of evaluating dietary intake and pave the way for personalized interventions for behavioral medicine.