We present a deep learning model for simultaneous range and direction of arrival estimation of frequency modulated continuous waveform radars using 1-bit analog-to-digital converters. Standard fast Fourier transform based processing of 1-bit signals can suffer from various signal distortions. To combat this, we train a neural network on vectorized covariance matrices of simulated targets and test it under realistic settings in terms of signal-to-noise ratio and computation. We also test under off-grid settings not seen during training and demonstrate that this type of model can indeed generalize. Furthermore, we present results from data collected from a 77 GHz automotive radar under low resolution settings and can accurately detect the position of a vehicle within the field of view.
Paper
Full text
Deep Learning Based Range and Doa Estimation using low Resolution FMCW Radars
Semantic Scholar · Engineering · 2021
Abstract
We present a deep learning model for simultaneous range and direction of arrival estimation of frequency modulated continuous waveform radars using 1-bit analog-to-digital converters. Standard fast Fourier transform based processing of 1-bit signals can suffer from various signal distortions. To combat this, we train a neural network on vectorized covariance matrices of simulated targets and test it under realistic settings in terms of signal-to-noise ratio and computation. We also test under off-grid settings not seen during training and demonstrate that this type of model can indeed generalize. Furthermore, we present results from data collected from a 77 GHz automotive radar under low resolution settings and can accurately detect the position of a vehicle within the field of view.