Interference Suppression for Radar Signal using 2D UNet based on Semantic Segmentation

The wide application of automotive radars significantly increases the possibility of mutual interference. Interference can lead to false detections such as ghost objects and missed detections, with serious threats to vehicle and pedestrian safety. In this paper, we build an end-to-end interference suppression model using 2D UNet. The UNet takes the input through the encoder with down-sampling to get a feature smaller than the initial data, and then inputs this feature into the decoder and reduces it to the clean signal, thus achieving interference suppression. By providing the network with clean data and interference-contaminated data, the network can be well trained to mitigate the interference artifacts. Experimental results show that the proposed scheme could achieve superior performance compared with traditional signal processing algorithms, in which the target peak was preserved and the signal-to-noise ratio (SINR) was significantly improved.