A Comparison of Convolutional Neural Networks for Low SNR Radar Classification of Drones
Holly Dale, Chris Baker, Michail Antoniou, Mohammed Jahangir, George Atkinson
Abstract
Reliable detection and tracking is required to ensure that drones are safely integrated into low altitude airspace. Radar provides a 24-hour, all-weather solution to this problem. However, the radar signatures of birds have a similar RCS to those of drones, thus a robust method of classification is needed to filter out non-drone targets and eliminate, or at least minimize to an acceptable level, false alarms. Convolutional neural networks (CNNs) have been shown to achieve high classification performance but results are only reported for high signal to noise ratio data -a luxury that is not always available to operational radar systems. In this paper, Gaussian noise is added to the test data to vary the signal to noise ratio (SNR) in order to investigate classifier robustness as a function of SNR in the context of drone classification. The performance of six CNN architectures previously established for computer vision applications are exploited and compared with each other to assess classification performance and robustness with network depth.