Detection-Guided Attention for Selective Target Classification Using Radar Micro-Doppler Spectrograms

Abstract

Multiple targets in the same radar micro-Doppler spectrogram, such as an uncrewed aerial vehicle (UAV) surrounded by flocking birds, can confuse classification algorithms. Without knowing which of the targets to classify, the decision is ambiguous. We propose to inform the classifier which targets to classify by encoding the detected target position as a separate channel. This instructs the convolutional neural network to pay attention to the selected target without removing context. We, therefore, enable the model to classify individual objects in multitarget spectrograms, paving the way for higher classification performance in complex environments. Different representations of the detection-guiding matrix are tested, and the approach is compared to alternative approaches, such as centering and cropping, and we show that it is superior in cases with multiple targets. The efficacy of the approach is demonstrated on synthetic multitarget spectrograms using multiple datasets.