Fully bayesian learning and spatial reasoning with flexible human sensor networks

This work considers the challenging problem of simultaneous modeling and fusion of 'soft data' generated by a network of 'human sensors' for spatial state estimation tasks, such as lost target search or large area surveillance. Human sensors can opportunistically provide useful information to constrain difficult state estimation problems, but are imperfect information sources whose reliability cannot be easily determined in advance. Formal observation likelihood models are derived for flexible sketch-based observations, but are found to lead to analytically intractable statistical dependencies between unknown sensor parameters and spatial states of interest that cannot adequately characterized by simple point estimates. Hierarchical Bayesian models and centralized inference strategies based on Gibbs sampling are proposed to address these issues, especially in cases of sparse, noisy, ambiguous and conflicting soft data. This leads to an automatic online calibration procedure for human sensor networks, as well as conservative spatial state posteriors that naturally account for model uncertainties. Experimental outdoor target search results with real spatial human sensor data (obtained via networked mobile graphical sketch interfaces) demonstrate the proposed methodology.