Machine Learning for Robust Automatic Target Recognition

Toyon Research Corporation and Dr. David Miller from The Pennsylvania State University propose to develop robust ATR algorithms for the classification of ground vehicles. Additionally, we propose to apply the same underlying technology to improve the robustness of feature-aided tracking (FAT). Our approach is based on recent advances in robust classifier design using semisupervised learning. That is, the classifier processes a set of data where some items have labels and some do not. Our approach can automatically recognize targets that are not part of the dataset used to train an ATR. Furthermore, it can make mappings between labeled data, such as synthetic signatures, and unlabeled measured signatures that are collected in the field. For FAT, our algorithms apply to both class-dependent and class-independent approaches and they allow for the development of new approaches which combine the benefits of both. In Phase I, we will focus on the application of our algorithms to high-range-resolution GMTI range profiles. We will train an ATR and evaluate its performance including the ability of the ATR to recognize targets that are not in the training set. We will finish with a final report and a plan for a Phase II effort.