Automotive Radar Interference Mitigation using a Convolutional Autoencoder
Jonas Fuchs, Anand Kumar Dubey, Maximilian Lübke, Robert Weigel, Fabian Lurz
Abstract
Automotive radar interference imposes big challenges on signal processing algorithms as it raises the noise floor and consequently lowers the detection probability. With limited frequency bands and increasing number of sensors per car, avoidance techniques such as frequency hopping or beamforming quickly become insufficient. Detect-and-repair strategies have been studied intensively for the automotive field, to reconstruct the affected signal samples. However depending on the type of interference, reconstruction of the time domain signals is a highly non-trivial task, which can affect following signal processing modules. In this work an autoencoder based convolutional neural network is proposed to perform image based denoising. Interference mitigation is phrased as a denoising task directly on the range-Doppler spectrum. The neural networks shows significant improvement with respect to signal-to-noise-plus-interference ratio in comparison to other state-of-the-art mitigation techniques, while better preserving phase information of the spectrum compared to other techniques.