Deep Learning Based Computationally Efficient Unrolling IAA for Direction-of-Arrival Estimation

We introduce a computationally efficient approach for direction-of-arrival (DOA) estimation in automotive radar systems using a single-snapshot. Classical subspace-based methods like MUSIC and ESPRIT may apply spatial smoothing on uniform linear array to create multiple snapshots for accurate DOA estimation. However, spatial smoothing has the drawback of reducing the array aperture and it is not feasible for sparse linear arrays. The existing single-snapshot-based methods like compressive sensing and iterative adaptation approach (IAA) have high computational costs and slow convergence times, which poses challenges for real-time implementations. While strides in optimization algorithms and hardware acceleration strategies propose plausible remedies to alleviate these constraints, enhancing their appropriateness for real-time use, the computational cost remains notably high. The recent deep learning-based DOA estimation methods have shown good performance in terms of inference time and estimation accuracy, but lack interpretability and generalization. To address these limitations, we propose an unrolling iterative adaptive approach (UAA) that unrolls the IAA algorithm into multiple deep neural network layers. The UAA network has better generalization and avoids the high computational costs associated with matrix inversions. Extensive numerical experiments show that the UAA network outperforms IAA in terms of inference time and estimation accuracy under different signal-to-noise ratio (SNR) scenarios.

I. Introduction

Automotive multiple-input multiple-output (MIMO) radars are an essential part of advanced driver assistance systems and self-driving cars, mainly because they are low cost, capable of sensing in bad weather, and unaffected by poor visibility conditions [1]–[6]. Frequency-modulated continuous-wave (FMCW) is commonly used in automotive radar systems with low-cost analog-to-digital converters (ADCs). The targets are separated in range-Doppler domains using two-dimensional fast Fourier transform (FFT), and a constant false alarm rate (CFAR) detector is used to select a subset of range-Doppler bins for direction-of-arrival (DOA) estimation through a third FFT. As a result, current automotive radar only provides sparse point clouds. To improve the angular resolution and generate high-resolution radar images, automotive radar can perform high-resolution DOA estimation for each range-Doppler bin to produce range-azimuth spectra imaging in bird's-eye view format [7]–[10].