Deep Learning-Based Segmentation for the Extraction of Micro-Doppler Signatures

Abstract

We present a method for extracting micro-Doppler signatures using a deep convolutional neural network that learns to identify and separate relevant micro-Doppler components from the background. A modified convolutional neural network (fully convolutional network) is trained end-to-end to perform dense predictions from the micro-Doppler signature at the input, generating a map with labels on a pixel level at the output. The network learns intermediate representations with the characteristic patterns of the micro-Doppler paths generated by individual scatterers and is capable of identifying and locating them in the time-frequency representation. The model trained on a simulated environment shows very good performance metrics even in noisy environments, and the experimental results with a continuous wave (CW) radar at 24 GHz indicates that the model can be applied to real scenarios. Moreover, the method scales properly to more complex signatures when several components are superimposed in the time-frequency representation, which indicates that this concept might represent a promising approach for interpreting complex micro-Doppler signatures.