Applications of Neural Networks for FMCW Automotive Radars |Techniques for Improving Detection and DoA Estimation
Marcio Lima de Oliveira is a PhD student in the department Computer Architecture Design and Test for Embedded Systems. (Co)Promotors are prof.dr.ir. M.J.G. Bekooij and dr.ir N. Alachiotis from the faculty Electrical Engineering, Mathematics & Computer Science.
The automotive industry has undergone a major transformation, with vehicles evolving from basic mechanical systems to advanced electromechanical machines packed with sensors such as 3D cameras, LiDAR and FMCW radars. Automotive radars have become essential to modern driver assistance systems, capable of providing accurate estimates of the speed, distance and angle of targets.
Over time, radar technology has evolved from their military origins to reliable, low-cost automotive components. However, the growing range of applications has presented new challenges, including the need for low latency, limited computational capacity, and antenna constraints. Traditional solutions have relied on handcrafted, analytical signal processing pipelines to solve these problems. Meanwhile, new machine learning techniques are generating increasing interest in radar applications.
While deep learning has transformed many fields, radar signal processing still relies on well-understood mathematics and physics. On the other hand, many of these traditional models rely on approximations and assumptions that may not always hold in real-world situations. Therefore, integrating deep learning with analytical signal processing might be a beneficial way to overcome some of the current limitations of the automotive radar pipeline while taking advantage of the strengths of both approaches.
In this thesis, we discuss automotive radar signal processing, focusing on the integration of deep learning techniques to complement the traditional analytical signal processing pipeline. Our underlying objective is to reduce the gap between deep learning and analytical techniques for automotive radars. This thesis addresses five key problems: (i) detecting moving objects in various scenarios and noise levels using range-Doppler maps, (ii) generating representative synthetic range-Doppler maps for training and evaluating other neural networks, (iii) estimating the direction of arrival of targets using a non-linear antenna array and a single snapshot, (iv) estimating the number of targets for multiple antennas, and (v) incorporating imperfections into our antenna models.
