FMCW Radar Signal Processing Pipeline for Aircraft Marshalling Signals Classification

We present a methodology for obtaining a low memory and low computational load for a radar signal processing pipeline of a Frequency-Modulated Continuous-Wave (FMCW) radar tailored to embedded systems. Our analysis covers the entire radar signal processing pipeline, including radar signal conditioning, feature extraction, and neural network models, including pruning and quantization. We focus on a complex dataset of aircraft marshalling signals captured from an ultra-low-power, single-input, single-output (SISO) FMCW radar. With our methodology, we achieve a new state-of-the-art classification accuracy of 71.4\%, marking a significant improvement of over six percentage points while maintaining a reduced memory footprint and minimal computing requirements. This advancement contributes to developing efficient radar signal processing systems for resource-constrained applications and represents a crucial step toward realizing highly efficient and accurate radar-based gesture recognition systems.