Lightweight power aware and scalable movement monitoring for wearable computers: a mining and recognition technique at the fingertip of sensors

Activity monitoring using Body Sensor Networks(BSN) has gained much attention from the scientific community due to its recreational and medical applications. Suggested techniques for activity monitoring face two major problem. First, systems have to be trained for the individual subjects due to the heterogeneity of the BSN data. While most solutions can address this problem on a small data set, they have no mechanics for automatic scaling of the solution as the data set increases. Second, the battery limitations of the BSN severely limit the maximum deployment time for the continuous monitoring. This problem is often solved by shifting some processing to the local sensor nodes to avoid a very heavy communication cost. However, little work has been done to optimize the sensing and processing cost of the action recognition. In this paper, we propose an action recognition approach based on the BSN repository. We show how the information of a large repository can be automatically used to customize the processing on sensor nodes based on a limited and automated training process. We also investigate the power cost of such a repository mining approach on the sensor nodes based on our implementation. To assess the power requirement, we define an energy model for data sensing and processing. We demonstrate the relationship between the activity recognition precision and the power consumption of the system during continuous action monitoring. We demonstrate the energy effectiveness of our approach with a classification accuracy constraint based on limited data repository.