Radar Interference Mitigation Using CNN Autoencoders with Spatial Attention

As part of the growth in the automotive industry, there is an increase in the number of radar sensors that are deployed today. This growth comes at the cost of a potential increase in interference emanating from neighboring radar sensors that are within range. Traditionally, signal processing techniques have been used to mitigate interference but deep learning methods have drawn significant attention in recent times. To this end, we propose Radar-SACAE (Radar with Spatial Attention and Convolutional Denoising Autoencoder). This a deep learning model that applies spatial attention to the current range-doppler map input and previous inputs. This is then passed through a convolutional autoencoder to achieve very interesting performance, compared to existing models, in terms of signal-to-interference and noise ratio and error vector magnitude. This optimal performance is achieved with a highly significant reduction in computational complexity compared to that of other deep learning approaches.