Spectral-Kurtosis and Image-Embedding Approach for Target Classification in Micro-Doppler Signatures

Abstract

Micro-Doppler signature represents the micromotion state of a target, and it is used in target recognition and classification technology. The micro-Doppler frequency appears as a transition of the Doppler frequency due to the rotation and vibration of an object. Thus, tracking and classifying targets with high recognition accuracy is possible. However, it is difficult to distinguish the types of targets when subdividing targets with the same micromotion or classifying different targets with similar velocities. In this study, we address the problem of classification of three different targets with similar speeds and segmentation of the same type of targets. A novel signature extraction procedure is developed to automatically recognize drone, bird, and human targets by exploiting the different micro-Doppler signatures exhibited by each target. The developed algorithm is based on a novel adaptation of the spectral kurtosis technique of the radar echoes reflected by the three target types. Further, image-embedding layers are used to classify the spectral kurtosis of objects with the same micromotion. We apply a ResNet34 deep neural network to micro-Doppler images to analyze its performance in classifying objects performing micro-movements on the collected bistatic radar data. The results demonstrate that the proposed method accurately differentiates the three targets and effectively classifies multiple targets with the same micromotion.