Energy-Efficient Human Activity Recognition on Smartphones: A Test-Cost Sensitive Approach

Human activity recognition is essential for providing services in the Internet of Things. Thanks to their ubiquity, sensing capability, and processing power, modern smartphones have become attractive devices for activity recognition. However, their limited battery capacity places a hurdle to exploit such sensing and processing power. While power is consumed in both the sensing and computation layers of the recognition process, power optimization in the latter layer has not been studied extensively enough. This paper strives towards energy-efficient activity recognition by focusing on the cost of feature extraction. To this end, the energy cost of extracting various features is examined and test-cost sensitive prediction models are employed to recognize activities from features. Experimental results reveal a considerable opportunity to conserve energy by awareness of the cost of feature extraction. With only a small sacrifice in prediction accuracy, the energy cost of computations can be reduced by a factor of three.

Keywords:

Internet of Things (IoT) , Human Activity Recognition (HAR) , Power-Aware Computing , Pervasive Computing , Test-Cost Sensitive Learning , Ambient Intelligence (AmI)