An Investigation of the Effects of Correlation in Sensor Fusion

This thesis takes the first step towards the creation of a synthetic classifier fusion-testing environment. The effects of data correlation on three classifier fusion techniques were examined. The three fusion methods tested were the ISOC fusion method (Haspert, 2000), the ROC "Within" Fusion method (Oxley and Bauer, 2002) and the simple use of a Probabilistic Neural Network (PNN) as a fusion tool. Test situations were developed to allow the examination of various levels of correlation both between and within feature streams. The effects of training a fusion ensemble on a common dataset versus an independent data set were also contrasted. Some incremental improvements to the ISOC procedure were discovered in this process.