Influence of Radar Signal Processing on Deep Learning-based Classification

Abstract

As radar technology becomes more readily available to researchers and users, it is thus being explored how to better process this data for real-time implementations. To process this radar data, the short time Fourier transform (STFT) has been implemented to then find the micro-Doppler spectrogram. When computing the STFT, there are parameters which can be adjusted to alter the size of the resulting micro-Doppler spectrogram. In this work, these parameters were adjusted to find the optimal representation of micro-Doppler radar returns of human activities, which were recorded using a 77 GHz Frequency Modulated Continuous Wave (FMCW) millimeter wave radar. To determine these optimal combinations, the resulting micro-Doppler spectrograms were used to train and test a Convolutional Autoencoder (CAE). The t-Distributed Stochastic Neighbor Embedding (t-SNE) and k-Nearest Neighbor Classification (kNN) were also utilized to find the nearest representations in a low-dimensional space of the spectrograms.