Single MLP-CFAR for a radar Doppler processor based on the ML criterion. Validation on real data

This paper tackles the evaluation of radar detectors with real data in a scenario composed by targets with unknown Doppler shift and sea clutter. A Neural Network-based Constant False Alarm Rate (CFAR) technique, NN-CFAR, is compared with reference detection schemes based on Doppler processors and conventional CFAR detectors. In these reference solutions, although CFAR techniques are designed for a desired false alarm rate, PFA, we prove that the final PFA rate is higher than the desired one. In this paper, a detection performance improvement is obtained with a detector that is a better approximation to the Neyman-Pearson detector based on the generalized Likelihood Ratio (selecting the maximum filter bank output), and uses a unique CFAR detector. Due to the non-linear nature of the maximum function, conventional CFAR detectors are not suitable. The improved detector is designed and applied to real data acquired by a coherent and pulsed radar system at X-band frequencies. Results prove that the NN-CFAR provides a higher probability of detection while fulfilling the PFA requirement.