Work Package 5 Research
SPHERE researchers are developing theoretical and empirical evaluation of static and structured classification models on structured datasets which should inform SPHERE of what models to use. Researchers can encode the structural nature of these problems into features, which:
• reduces training time significantly
• reduces number of parameters to estimate
• increases interpretability of models
SPHERE would like to use groups of sensors together to create classifiers that are robust in a multi-resident environment. Given a new house, is it possible to automatically learn the topology of the sensors simply from sensor activations? SPHERE has shown that through the application of standard signal processing and information theoretic approaches, it is possible to automatically learn about the topology of the house. Researchers have demonstrated this result on several datasets, showing that the method is robust to the layout of the house and the number of residents in the house.
In parallel, SPHERE researchers at the University of Reading are aiming to provide access to fine details of the user’s motion for applications such as rehabilitation. To do so, an innovative method is being developed to recover the 3D body posture with a minimal number of body worn inertial sensors.