Heteroscedastic Gaussian Process Model for Received Signal Strength Based Device-Free Localization

Received signal strength (RSS) based passive localization approaches measure human-induced changes in the electromagnetic field to localize and track people. Bayesian estimation methods have been widely utilized to solve the problem, mainly because of their convenience in representing uncertainties in the models and in modeling physical randomness. The localization performance is significantly influenced by the measurement model that describes the electromagnetic field changes as a function of the location of the target, and a wide variety of empirical and analytical models have been proposed. Common to these models is that the measurement noise is assumed homoscedastic, that is, the measurement noise is constant. In this paper, the measurement noise is assumed to depend on the location of the target, and a novel heteroscedastic Gaussian process model for RSS-based device-free localization and tracking (DFLT) is proposed. In addition, algorithms to train the model parameters and solve the RSS-based DFLT problem are presented. The models and tracking algorithms are evaluated using experiments conducted in an open-space indoor environment and in a fully furnished downtown residential apartment. The results imply that the proposed approach can decrease the localization error with respect to the benchmark RSS models and that real-time sub-decimeter tracking accuracy can be achieved in both environments.