Devices like smartwatches are often used to track physical activity, helping users meet their exercise goals. This tracking relies on accurate inference of the physical activity being performed from data available to the device. These data may look similar between activities which the user would prefer to distinguish, making quick and accurate activity classification challenging.

In this project, we build models capable of classifying physical activities based on data accessible by consumer smart devices and test their accuracy against each-other. Once trained, these models can run live on consumer devices such as smart watches, classifying periods of activity and displaying progress towards the user's movement goals. In addition to consumer use, data classified by these models could have clinical applications to long-term medical tests such as Holter monitors.