Learning based multiple input multiple output radar imaging

Multiple Input Multiple Output (MIMO) radars, with their unique antenna configuration,offer superior angular resolution by forming a “virtual array” withfewer antennas compared to phased-array radars. Through digital beamformingtechniques, MIMO radars can generate 3D radar images of their surroundingswithout the need for mechanical rotation. Reconstructing an unknown field suchas range/azimuth profile or full radar reflectivity image from radar returns, whichare both noisy and band-limited, presents a challenging and ill-posed inverse problem.Conventional reconstruction methods typically involve inversion of the measurementsthrough a fully known mathematical measurement model that mimicsthe underlying physics of measurements. However, conventional approaches oftenrely on a measurement model and model-based inversion techniques may not fullyleverage the prior knowledge of the unknown field being reconstructed when suchinformation is known. To incorporate prior information of the radar data into thereconstruction process, regularizers can be employed to promote specific spatialpatterns within the radar data. Nevertheless, these regularizers often fall short in effectively capturing the intricate spatial patterns within the field data being reconstructed,or they may not readily allow for analytical minimization of the costfunction. Recently, the Alternating Direction Method of Multipliers (ADMM)framework has emerged as a means to provide a way of decoupling the modelinversion from the regularization of the priors, enabling the incorporation of anydesired regularizer into the inversion process in a plug-and-play (PnP) fashion. Inthis thesis, we implement the ADMM framework to address the radar range profileand image reconstruction problems where we propose to employ deep learning networksas a regularization method for enhancing the quality of the inversion processwhich usually suffers from the ill-posed nature of the problem. We demonstrate theefficacy of deep learning networks as a regularization method within the ADMMframework via our simulation results for the inverse problems in MIMO radarcontext. We assess the performance of the ADMM framework employing CNNand U-Net architectures as a regularizer for radar range profile and radar imagereconstruction problems, respectively. We conduct a comparative analysis againstalternative methods under different measurement scenarios. Notably, among themethods under investigation, ADMM with deep learning networks used as a regularizerstands out as the most successful method for both radar range profile andimage reconstruction problems in terms of reconstruction error. Furthermore, westudy the problem of target angle estimation in MIMO radars, proposing the use ofartificial neural networks for this purpose. Through comprehensive experiments,we show the potential of our proposed method across a range of measurement scenarios,verifying its effectiveness especially in the presence of measurement