Power-Accuracy Tradeoffs in Human Activity Transition Detection

Wearable, mobile computing platforms are envisioned to be used in out-patient monitoring and care. These systems continuously perform signal filtering, transformations, and classification, which are quite compute intensive, and quickly drain the system energy. The design space of these human activity sensors is large

and includes the choice of sampling frequency, feature detection algorithm, length of the window of transition detection etc., and all these choices fundamentally trade-off power/performance for accuracy of detection. In this work, 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. For instance, we find that the x-axis of our signal, which was oriented to be 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.