Structured synthesis and compression of semantic human sensor models for Bayesian estimation

We consider the problem of fusing human-generated semantic `soft sensor' data with conventional `hard sensor' data to augment Bayesian state estimators. This requires modeling semantic soft data via generalized continuous-to-discrete softmax likelihood functions, which can theoretically model semantic descriptions of any dynamic state space. This paper addresses two important related issues for deploying these models in practical applications. First, a general solution to the data-free likelihood synthesis problem is provided. This allows for easy embedding of contextual constraints and other relevant a priori information within generalized softmax models, without resorting to expensive non-convex optimization procedures for parameter estimation with sparse data. This result is then used to derive strategies for combining multiple semantic human observation models into `compressed' likelihood functions for fast batch data fusion. The proposed methods are demonstrated on a human-robot target search application.