SmartSen: smart sensing for enhancing real-time activity recognition in phone-based interactive CPS

Abstract

Using mobile devices to enhance users interaction with CPS has many potential benefits for healthcare applications. Recent research has looked at how to recognize human activities using smartphones, to indicate health status. But little attention has been paid to automatically identify the activity habits of individuals in real-time. Of course, the energy constraint in smartphones must be considered during the design of such cyber-physical recognition systems. In this paper, we propose a prediction-based smart sensing strategy that is energy efficient and works in real-time. By making use of the temporal correlation property of real-world phenomena, an adaptive k-order Markov chain based prediction algorithm is proposed to avoid continuous sensing so that significant energy savings can be achieved. The prediction results are analyzed online in real-time, to ensure that the system can track an individual's behavior pattern and provide timely response to changes in behavior. Real-world experiments using our prototype show that such recognition oriented CPS systems can not only achieve energy savings, but also converge to steady state with high individual recognition accuracy, in a real-time manner.