Contemporary smart cell phones contain a tri-axial accelerometer, which makes it possible to measure the acceleration of a smart phone user in all three spatial dimensions. Modern data mining methods permit the construction of an activity recognition classifier from this acceleration data, so that a user's physical activity (e.g., walking, jogging, standing, etc.) can be automatically inferred from the accelerometer values. This research project will build an activity recognition system that can be deployed as a downloadable cell phone application. Users will then be able to access a description of their activities via a web interface and can use this information to monitor and change their behavior. Thus this research can be used to address the many health-related problems that result from physical inactivity. It can also assist with other health-related problem, such as falling in the elderly, by detecting falls and providing automatic notification to caregivers. The data generated via the cell phone activity recognition system will also enable large scale epidemiological studies of activity levels and health that, due to prohibitive costs, were not previously possible.
Building a successful activity recognition system will require addressing many technical challenges. This research will require improved feature construction methods for transforming time-series data into representations suitable for example-oriented classification algorithms. It will also require the evaluation of alternate architectures for performing wide-scale real-time data mining, in order to determine how much of the activity recognition work should be performed on the client (i.e., smart phone) versus a centralized server. The adequacy of smart cell phones for performing reliable activity recognition will also be evaluated given the many constraints (e.g., limited battery life) imposed by these devices.
