Detecting drones with radars and convolutional networks based on micro-Doppler signatures

The detection of drones using radars is a problem of great importance due to the wide proliferation of drones that are being used in a variety of applications. In this paper, we propose a novel approach to convolutional neural network (CNN)-based drone detection using radar micro-Doppler signatures. The CNNs are trained on micro-Doppler signatures obtained from short-time Fourier transform spectrograms of the time-series data of the radar reflections from the drones. In particular, we investigate the binary classification of drones versus noise using both simulated data and real data taken from rapidly-manoeuvring drones. First, we train a CNN to detect and classify drones using simulated data based on the Martin-Mulgrew (MM) model. We find that at a 10-decibel signal-to-noise ratio, this CNN performs with an F1 score greater than 0.8. Furthermore, we apply transfer learning on the trained model to adapt it to real data. We show that this use of transfer learning improves the results over a standalone model trained solely on real data by 0.075 F1 score points.