Neurally Augmented Deep Unfolding for Automotive Radar Interference Mitigation

Abstract

The proliferation of active radar sensors deployed in vehicles has increased the need for mitigating automotive radar-to-radar interference. While simple avoidance and mitigation methods are still effective today, the expected crowded spectrum allocations pose new challenges that likely require more sophisticated techniques. In particular, interference mitigation methods that can handle significant levels of radar signal corruption are required. To this end, we propose neurally augmented analytically learned fast iterative shrinkage thresholding algorithm (NA-ALFISTA), which is a neural network-based solution for reconstructing time-domain radar signals by leveraging sparsity in the range-Doppler map (RDM). The neural augmentation network is deployed as a single gated recurrent unit (GRU) cell that captures the radar signal statistics along the unfolded layers of fast-iterative shrinkage thresholding algorithm (FISTA)-based sparse recovery, which significantly boosts the convergence rate. It estimates the next layer’s parameters necessary in ALFISTA based on the previous layer’s output. The proposed method is compared to state-of-the-art detect-and-repair methods and source separation methods in simulated data and real-world measurements.