Determining intent using hard/soft data and Gaussian process classifiers

Modern applications of data fusion are rarely starved of data but they look more challenging because the data can be diverse (hard and soft), uncertain and ambiguous, and often swamped by irrelevant detail. This paper presents a mathematical framework for dealing with these issues. It manages diverse data by representing it in the common format of a kernel matrix. Uncertainty is managed by interpreting the kernels as the covariance matrices of Bayesian Gaussian Processes. Finally, irrelevant detail is managed by automatically detecting the relevant (or, conversely, the irrelevant) components within a multi-source dataset. The framework is illustrated by applying it to a synthetic vehicle-borne Improvised Explosive Device intent recognition scenario. The results provide a proof-of-principle and encourage future work to develop practical implementations of algorithms in support of sense making and intelligence analysis.