Eating Detection Project


This project uses a wrist-worn device to track wrist motion all day and detect periods of time in which the person wearing the device is eating. Our current work uses the The Shimmer 3 device, but we have also processed data collected by the Actigraph gt9x device, Apple iPhones, and others.

Detecting periods of time of eating Shimmer 3 ActiGraph gt9x


The below figure shows a plot of wrist motion energy (total amount of linear acceleration) of a person for an entire day (7:30 AM to 7:30 PM). The subject kept a record of periods of time of eating, which are indicated by arrows along with the names of the meal (e.g. breakfast). We discovered that eating periods tend to have lower values of wrist motion energy surrounded by peaks of higher wrist motion energy. The higher peaks are typically due to meal preparation and cleanup in which the hands and wrists are moving a lot more than during actual consumption. Our algorithm uses peaks to segment the data and then calculates additional features between peaks to classify the periods of time.

Example output

The below figure shows an example of our classifier output. The y-axis is wrist motion energy (total amount of linear acceleration) and x-axis is time of day (10:30 AM to midnight). The subject's self-reported times of eating for lunch and dinner are indicated. Arrows above wrist motion energy indicate the boundaries of time our algorithm used to segment the data. Classifier output (eating, other, rest or walking) is indicated below the wrist motion energy.


We have collected a very large data set (351 people, 1 day each) of recordings like the one shown above. See the Clemson All-day Dataset (CAD) webpage for information and download.

Papers about this project:

Eating Detection Page / Clemson /