A 1D Cascaded Denoising and Classification Framework for Micro-Doppler-Based Radar Target Recognition

Micro-Doppler signatures play a crucial role in capturing target features for the radar classification task, and the time–frequency distribution method is widely used to represent micro-Doppler signatures in many applications including human activities, ground moving target identification, and different types of drones distinguishing. However, most existing studies that utilize radar micro-Doppler spectrograms often require extended observation times to effectively represent the cyclostationarity and periodic modulation of radar signals to achieve promising classification results. In addition, the presence of noise in real-world environments poses challenges by generating weak micro-Doppler features and a low signal-to-noise ratio (SNR), leading to a significant decline in classification accuracy. In this paper, we present a novel one-dimensional (1D) denoising and classification cascaded framework designed for low-resolution radar targets using a micro-Doppler spectrum. This framework provides an effective signal-based solution for feature extraction and recognition from the single-frame micro-Doppler spectrum in a conventional pulsed radar system, which boasts high real-time efficiency and low computation requirements under conditions of low resolution and a short dwell time. Specifically, the proposed framework is implemented using two cascaded subnetworks: Firstly, for radar micro-Doppler spectrum denoising, we propose an improved 1D DnCNN subnetwork to enhance noisy or weak micro-Doppler signatures. Secondly, an AlexNet subnetwork is cascaded for the classification task, and the joint loss is calculated to update the denoising subnetwork and assist with optimal classification performance. We have conducted a comprehensive set of experiments using six types of targets with a ground surveillance radar system to demonstrate the denoising and classification performance of the proposed cascaded framework, which shows significant improvement over separate training of denoising and classification models.