Energy Efficient On-Sensor Processing for Online Activity Recognition.

In sensor-based online activity recognition, the communication of sensor samples at high data rates has a great impact on the energy consumptions of wearables. In our work we investigate the idea of calculating data reducing stages of activity recognition systems on wireless sensor nodes in order to reduce the amount of transmitted data and thus the overall energy consumption. In our experiments, this approach could reduce the energy consumption of a wireless sensor node by up to 27%. Since the benefit of this approach highly depends on design parameters of the activity recognition, we introduce an energy trade-off model for wireless sensor nodes to estimate energy-savings of application specific configurations at design time. By calibrating this model for our wireless sensor node, we could achieve an accuracy of more than 99% in our experiments.