Energy saving for activity recognition through sensor selection, scheduling and sampling rate tuning

Activity recognition in pervasive environments relies often on many different sensors. Low-level data streams from these sensors are combined at the decision level at a sink. A key challenge is to efficiently and with high accuracy recognise the objects activities based on low-level sensor data. There is a trade-off between obtaining the required accuracy of activity recognition and the energy consumption.We formalise the problem of minimising energy consumption to recognise activities from the sensed data with application required accuracy. In contrast to existing work, we modulate this problem to consider not only the number of sensors to be used, but also the number of samples to be collected from them. This approach saves more energy and extends the network lifetime more than traditional approaches. Evaluation results with a publicly available dataset demonstrate that this approach extents the network lifetime by about 25% compared to traditional ones. Furthermore, our solution consumes a rather small amount of computing resources and can therefore be used in real-world settings.