Abstract
Non-cooperative target identification (NCTI) has become more and more important during the last years. So it is essential to develop robust classification schemes, which can be applied reliably in military operations. The major problems are the clutter (which is not present in air), the unknown target orientation (aspect and elevation angle), the high variability of the target scatterers and the possible multiple variants of the target. In this paper we concentrate on analysis and comparison of classification rates for target identification, the former detection process of the targets in the observed SAR scene will not be discussed. We focus on SAR images of single separated stationary and moving ground targets. To be able to compare our classification results with results published in literature we start our evaluation with the public MSTAR dataset, which is used since many years for ATR evaluation and identification. In addition to this we worked with a second dataset from another field measurement campaign which was performed by QinetiQ, UK. We quantify the influence of some fundamental key aspects on the classification rate. These subjects are target centering (using center of mass algorithms), image segmentation (target, clutter, target shadow), different applied classifiers (nearest neighbour classifier, support vector machines, SVM), the image resolution and the target orientation (knowledge of the aspect angle). Additionally we analyse the influence of clutter and target shadow on the classification rate when using again the MSTAR and QinetiQ datasets. For this purpose the data were segmented in a part containing the target (or target shadow) and another part containing only the surrounding clutter. For the QinetiQ dataset the classification rates drop from 79% to 70% when only the target (separated from the clutter) was used as a feature. The MSTAR dataset shows similar results. Thus obviously the identical background of each of the targets in the test and reference database contributed a lot to the reported high classification rates. That again shows that realistic expressions on the capabilities of target classification can only be based on independent test and training data. By introducing the image clutter content, ICC, we quantify the influence of the separated clutter on the classification rate. Furthermore the target shadow can be used for additional information dependent on the depression angle. Finally we come to the conclusion that the main work is not only choosing and applying the classifier, but concentrate on the data collection (that means good data quality), preprocessing and feature extraction processes.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have