Leveraging Machine Learning in CFAR Detectors

Abstract

Within the field of radar detection, a property called constant false alarm rate (CFAR) is of great importance. This property is used when, for example, creating generalized likelihood ratio test(GLRT)-based CFAR adaptive detectors. By using the statistics of these detectors, the authors of the article CFAR Feature Plane: A Novel Framework for the Analysis and Design of Radar Detectors have mapped radar data to a two dimensional feature space, called CFAR feature plane. During this work, this mapping chain was used to map data of high dimension to the plane, making it possible to compare clusters that forms in the feature plane. For larger steering vectors as input data (dimension of evaluated cell), there were less characteristics present in terms of amplitude and rotation between clusters of targets and clutter.

Furthermore, multiple estimations of the covariance matrix were used, both in terms of the amount of samples, but also together with regularization techniques, such as diagonal loading and fast maximum likelihood estimation. What could be seen was that the greater amount of data used for estimation, or by utilizing one of the regularization techniques, the more distinct clusters where formed in the CFAR feature plane, making the final classification easier. The main goal of this project was to implement different machine learning algorithms, trained in the feature plane, to investigate if it was possible to get a more robust detector in terms of mismatched targets, than the traditional model based ones, such as Kelly’s detector.

Characteristics for the clusters, how each cluster is distributed in feature space, directly affect detector performance. The larger the steering vector, the more intersection between clusters, causing the trained machine learning detectors to adapt to the behavior of a Kelly’s detector, performing relatively well. As for the smaller sizes of input data, it is possible to create machine learning detectors that performs better than the traditional ones, both in terms of perfect matched and mismatched targets. Such algorithms are multilayered perceptrons and symbolic classifiers.