Radar Target Detection with CNN

Target detection is a fundamental radar application that is traditionally carried out by Constant False Alarm Rate (CFAR) detectors. This paper proposes a Convolutional Neural Network (CNN) based detector (RadCNN) to replace the standard CFAR detectors for a typical pulsed Doppler radar. RadCNN takes patches of the range-Doppler ambiguity function as input and returns detection status for the input patch. A radar simulator is developed for data generation with desired noise and clutter scenarios. RadCNN is compared against Cell-Averaging (CA), Smallest of Cell Averaging (SOCA), Greatest of Cell Averaging (GOCA), Ordered Statistics (OS) CFAR and similar state of the art detectors in the literature. The comparison is done for a variety of scenarios including multiple targets, thermal noise and clutter at different Signal to Noise Ratios (SNR) and Clutter to Noise Ratios (CNR). It is shown that RadCNN improves the performance of CFAR for low SNR and exhibits four orders of magnitude less computational complexity than the similar state of the art and realizable in real-time applications.