Micro-Doppler based Deep Learning approaches for radar applications

Technology is advancing more and more rapidly in terms of hardware like radars and their capacities to detect smallest signals and also in software like deep neural networks and their ability to translate hidden patterns of radar signals into meaningful representations. In our work, we focus the context of deep learning techniques and investigate the performance of different architectures of convolutional neural networks (CNNs) as well as transformers. The performance was compared on Micro Doppler Automatic Target Recognition ATR, which includes smaller objects like birds and drones which can only be detected by their Micro Doppler characteristics in a sufficient resolution. These state-of-the-art architectures were evaluated based on the number of parameters they used, as well as their performance on the task. By doing so we performed an architecture search and identified the best deep learning model for a company based hard to fit dataset.