Power/Accuracy Tradeoffs for Home Stroke Rehabilitation
Wearable, mobile computing platforms are envisioned to be used in out-patient monitoring and care. These systems continuously perform computationally intensive operations, quickly draining the system energy. While device hardware provides figurative nobs that can adjust the energy efficiency of a device, system software actually turns the nobs. We research methods to make software that recognizes activities of daily living smarter (i.e. more energy efficient) by using knowledge of human movement dynamics and behavioral patterns. Our goal is to create a software-optimized wearable computer that is as small, light, and unobtrusive as a wrist-watch, have a battery life of a week or more, and still be accurate in classifying human activity.
As a starting point, we look at the design space of human activity sensing, which is large and includes sampling frequency, feature detection algorithms, length of the window of transition detection, sleep strategies etc., and all these choices fundamentally trade-off power/performance for accuracy of detection. We explore this design space, and make several interesting conclusions that can be used as rules of thumb for quick, yet power-efficient designs of such systems. We show that the x-axis of our signal, which was oriented parallel to the forearm, is the most important signal to be monitored, for our set of hand activities. Our experimental results show that by carefully choosing system design parameters, there is considerable (5X) scope of improving the performance/power of the system, for minimal (5%) loss in accuracy.