Advancing High-Resolution and Efficient Automotive Radar Imaging through Domain-Informed 1D Deep Learning
Ruxin Zheng (University of Alabama)
Shunqiao Sun (University of Alabama)
Hongshan Liu (University of Alabama)
Holger Caesar (TU Delft - Mechanical Engineering)
Honglei Chen (MathWorks, Natick)
Jian Li (University of Florida)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Millimeter-wave (mmWave) radars are critical for autonomous vehicles' perception tasks, offering reliable performance in adverse weather conditions. However, their application is often hindered by insufficient spatial resolution for detailed semantic scene interpretation. Traditional super-resolution methods derived from optical imaging fail to accommodate the unique properties of radar signals. Addressing this, our study redefines radar imaging superresolution as a one-dimensional (1D) signal super-resolution spectra estimation problem, leveraging domain-specific insights to innovate data normalization and introduce a domain-informed signal-tonoise ratio (SNR)-guided loss function. Our custom deep learning network, tailored for automotive radar imaging, achieves substantial improvements in parameter efficiency, and inference speed while enhancing image quality and resolution. Comprehensive tests demonstrate that our SR-SPECNet establishes a new standard for high-resolution radar range-azimuth imaging, surpassing previous methods. Source code and new radar dataset will be made publicly available at https://github.com/ruxinzh/SR DOA.