Avoid Touching Your Face: A Hand-to-face 3D Motion Dataset (COVID-away) and Trained Models for Smartwatches

World Health Organisation (WHO) advises that humans must try to avoid touching their eye, nose and mouth, which is an effective way to stop the spread of viral diseases. This has become even more prominent with the widespread coronavirus (COVID-19), resulting in a global pandemic. However, we humans on average touch our face (eye, nose and mouth) 10-20 times an hour [22] [12], which is often the primary source [15] of getting infected by a variety of viral infections including seasonal Influenza, Coronavirus, Swine flu, Ebola virus, etc. Touching our face all day long is a quirk of human nature [13] and it is extremely difficult to train people to avoid touching their face. However, wearable devices and technology can help to continuously monitor our movements and trigger a timely event reminding people to avoid touching their face. In this work, we have collected a hand-to-face multi-sensor 3D motion dataset and named it COVID-away dataset. Using our dataset, we trained models that can continuously monitor human arm/hand movement using a wearable device and trigger a timely notification (e.g. vibration) to warn the device users when their hands are moved (unintentionally) towards their face. Our trained COVID-away models can be easily integrated into an app for smartwatches or fitness bands. Our evaluation shows that the Minimum Covariance Determinant (MCD) model produces the highest F1-score (0.93) using just the smartwatch’s accelerometer data (39 features). Both the dataset and trained models are openly available on the Web at https://github.com/bharathsudharsan/COVID-away.