Recent investigations into multi-sensor fusion have yielded a variety of data fusion algorithms. Some fuse imagery from
multiple sensors at the pixel level, while others fuse outputs of detection algorithms-such as radar prescreeners or
hyperspectral anomaly detectors-at the feature level. Many of the feature-level fusion algorithms build upon the
foundation of Bayesian probability, and they assign probability to the event that a certain feature value is due either to a
target or to clutter. A few of the feature-level fusion algorithms, however, exploit tools developed within the framework
of the Dempster-Shafer (DS) theory of evidence. In these formulations some of the probability can be assigned to a third
hypothesis representing uncertainty, and the algorithm developer must specify an uncertainty function that maps feature
values to probability "mass" for this third hypothesis. Unfortunately, the DS paradigm lacks a standard method for
assignment of mass to the "don't know" hypothesis for a particular input feature.
In this paper we define a feature-level DS fusion algorithm and determine a method for specifying its uncertainty
function. We begin by developing and describing a measure of performance based on the area under the receiver
operating characteristic (ROC) curve. We then incorporate this measure of performance into a training procedure that
exploits the dynamics of the particle swarm and is capable of discovering locally optimal uncertainty functions. We
exercise the training algorithm using simulated data, and analyze the performance of its hypothesized optimal
uncertainty function. Next we apply the newly developed training techniques to data produced by separate prescreener
algorithms operating on measured Hyperspectral Imager (HSI) and synthetic aperture radar (SAR) data from the same
scene. Finally, we quantify the performance of the entire DS fusion procedure using ROC curves.
|