Deep Learning Based Radar Target Classification Using Micro-Doppler Features

Abstract

Demand for radar automatic target recognition is ever increasing owing to the extensive employment of radar sensors in urban scenarios and a drastic increase in the number of radar targets, especially drones and UAVs. Micro-Doppler signatures, resulting from the micro-motion dynamics of targets, have emerged as a key distinctive feature for radar automatic target recognition. This paper addresses the problem of radar target recognition based on deep learning and micro-Doppler signatures of targets. The choice of MobileNetV2 deep Convolutional Neural Network based classification on spectrogram images of the targets, has made the system more suitable for system implementation on embedded devices such as Raspberry Pi. Second important contribution of this paper is the augmentation of an extensive and diverse training dataset having five classes ultimately, for the testing of radar automatic target recognition, since few such datasets are available in the open literature. The dataset is developed using a W-band Frequency Modulated Continuous Wave radar. After training the model on the diverse training dataset, validation and test accuracies of 98.67% and 99% respectively, are achieved.