Efficient SVD techniques for micro-Doppler signatures processing in mmWave radar systems

Abstract

This thesis investigates the extraction of micro-Doppler signatures in frequency-modulated continuous wave (FMCW) radar data under challenging scenarios, such as detecting hidden pedestrian targets while the radar is operating in the presence of interference from other FMCW radars. Reliable detection of micro-Doppler (MD) signatures is essential for applications such as pedestrian recognition in autonomous vehicle systems. However, challenges arise due to interference from other FMCW radars and reflections caused by environmental obstructions, such as parked vehicles and cars ahead on the road. To address this, we first show that short-time Fourier transform (STFT), which is commonly used for MD extraction, is not suitable under challenging scenarios. We show that singular value decomposition (SVD) of a matrix of in-phase (I) and quadrature (Q) complex radar samples can be used for MD extraction under challenging scenarios. However, for MD extraction, SVD needs to be performed on a large IQ sample matrix, which can be computationally intensive for radar data processing system on chips (SoCs). We study various computationally efficient SVD methods, such as incremental singular value decomposition (ISVD) and randomized singular value decomposition (RSVD). We applied ISVD and RSVD to separate pedestrian movements from other targets and radar interference signals. The effectiveness of various SVD techniques was evaluated based on their ability to preserve micro-Doppler features while minimizing signal distortion. Results indicate that ISVD and RSVD successfully retain essential micro-Doppler characteristics despite interference from another radar and signal reflections from other targets such as a car. Additionally, computational efficiency and reconstruction error were compared for the different SVD methods, revealing trade-offs between the various approaches. We also provide an implementation architecture for real-time SVD processing as a programmable hardware accelerator. We also make the case that for real-time radar SoCs, optimizations relating to SVD processing are needed to meet the latency deadlines of real-time system. Our findings suggest that both RSVD and ISVD methods offer promising solutions for FMCW-radar-based MD extraction analysis under challenging scenarios. Future direction of this work can focus on optimizing these techniques for real-time radar SoC processing application and extending them to more complex environments.