PITA Fiscal Year 2009 Projects - Information and Systems Technology

Improvements in Activity Recognition and Monitoring of User Activities

Principal Investigators: Asim Smailagic, Dan Siewiorek, Marvin Sirbu

The BodyMedia armband has proprietary algorithms for taking accelerometer data and predicting the amount of calories consumed. Currently these algorithms do not record physical activity. Furthermore they have not been tuned to preserve battery power. Our machine learning algorithms, developed on the eWatch, have been optimized to conserve energy without compromising accuracy of classification. We will port our machine learning algorithms to the BodyMedia armband to calibrate their accuracy on activity recognition. We will identify the tradeoff of battery power versus accuracy by varying sample frequency, percentage of sampling points, and aggressive sleep cycles based on predicting activity transitions from prior data. By integrating the eWatch algorithms with BodyMedia algorithms, we can optimize both for the best accuracy for the least amount of power consumption. Furthermore BodyMedia will be able to offer new services such as generating reports on the amount of time each day the user is engaged in each activity (running, walking, sitting, standing, etc). This project provides an opportunity for knowledge and technology transfer that would be beneficial for BodyMedia in their further applications in healthcare and activity monitoring.