User Movements Forecasting by Reservoir Computing Using Signal Streams Produced by Mote-Class Sensors

Real-time, indoor user localization, although limited to the current user position, is of great practical importance in many Ambient Assisted Living (AAL) applications. Moreover, an accurate prediction of the user next position (even with a short advice) may open a number of new AAL applications that could timely provide the right services in the right place even before the user request them. However, the problem of forecasting the user position is complicated due to the intrinsic difficulty of localization in indoor environments, and to the fact that different paths of the user may intersect at a given point, but they may end in different places. We tackle with this problem by modeling the localization information stream obtained from a Wireless Sensor Network (WSN) using Recurrent Neural Networks implemented as efficient Echo State Networks (ESNs), within the Reservoir Computing paradigm. In particular, we have set up an experimental test-bed in which the WSN produces localization information of a user that moves along a number of different paths, and in which the ESN collects localization information to predict a future position of the user at some given mark points. Our results show that, with an appropriate configuration of the ESN, the system reaches a good accuracy of the prediction also with a small WSN, and that the accuracy scales well with the WSN size. Furthermore, the accuracy is also reasonably robust to variations in the deployment of the sensors. For these reasons our solution can be configured to meet the desired trade-off between cost and accuracy.