Fully Bayesian Human-Machine Data Fusion for Robust Dynamic Target Surveillance and Characterization
Jeremy Muesing, Luke Burks, Michael L. Iuzzolino, Danielle Albers Szafir, Nisar R. Ahmed · AIAA Scitech 2019 Forum · 2019
This work examines the problem of rigorously characterizing and fusing observations provided by human operators with probabilistic information extracted by an automated data fusion system, in the context of dynamic multi-target track characterization for large-scale surveillance. This task is characterized by three major challenges. Firstly, stand-alone human operator observation errors are difficult to parameterize and calibrate a priori. Secondly, successive single target observations from one operator are, in general, not conditionally independent of one another. Finally, the decision of when a human operator should assist the automation depends on the observed operator error characteristics as well as the uncertainty of the automation's track characterizations, all of which must be calculated online. A new hierarchical fully Bayesian probabilistic model is developed to explicitly account both for uncertainties in `human sensor' quality and Markovian conditional dependencies in successive target characterization reports. This model is used to perform online Bayesian inference via Gibbs sampling to simultaneously update the data fusion system's knowledge of human sensor characteristics and target type probabilities. The probabilistic model also allows for online Value of Information assessments to automatically query help from the human operator. Practical methods for approximating high-dimensional human sensor parameter posterior distributions via Dirichlet pdf moment-matching and parameter-tying are also developed, and shown to provide significant computational speedups for little/no noticeable information loss. These methods also demonstrate scalability with an increase in possible target types. Results with different combinations of simulated operator profiles and automated fusion system target characterization baselines show that fully Bayesian fusion can simultaneously improve machine-based characterization performance with variable human operator profiles, while accounting for uncertain operator characteristics.