Multisensor target seeker - classification of ground targets using datafusion

This part of the project MUMS (Multi Sensor Target Seeker Demonstrator) deals with classification and datafusion of military target models in a simulated scene. A multisensor target seeker flying at constant height 300 m approaches one of six different combat vehicles at an initial distance of 8 km. The seeker is a two-sensor combination of IR (8-9 micro m) and HRR radar (16GHz). To enable classification, features have been extracted from the segmented images and radar profiles of the targets. The IR features have been pre-processed using canonical diskriminant analysis (enhances class separability) and independent component analysis (ICA). The dimension-reduced features have been fed to a number of different classification algorithms, such as: Bayes method using separable or multiple probability densities, Gauss mixture densities, Parzen, k-nearest neighbours, artificial networks (RPROP = Resilient Backpropagation) and Support vector machines (SVM). Radar targets have been classified using RPROP, SVM and correlation operating directly on the distance profiles. The results show that targets are classified almost 100 % correct already at 1700 m distance provided both IR and radar are used and their fused decisions are combined over several distances by majority voting. Using IR or radar only, correct classification with voting drops to about 65 % and 80 %, respectively. These results are computed using the same classifier, RPROP, for both sensors. Other combinations of classifiers (may be different for IR and radar) give similar results.