UAV Detection in Strong Micro-Doppler Clutter Using Denoising Autoencoders and Convolutional Neural Networks

Radar detection of unmanned aerial vehicles (UAVs) in urban environments is a challenging problem. Conventional micro-Doppler-based classification algorithms fail in the presence of strong, closely spaced clutter. Algorithms that can isolate radar UAV signatures are therefore critical to next-generation urban airspace monitoring systems. This paper presents a denoising autoencoder (DAE) to extract UAV signatures in the presence of other micro-Doppler sources such as a rotating fan and human motion (walking and waving). The isolated signal is then detected with a convolutional neural network (CNN). Laboratory experiments with a millimeter-wave radar show that the detection accuracy of the CNN falls from 98.75% for a single target to 66.7% when UAV, human, and fan signatures are simultaneously present. However, processing the micro-Doppler spectrogram with the DAE prior to CNN detection results in near perfect results with only minor CNN fine-tuning needed to handle data scaling changes introduced by the DAE.