CMxUNet: An Efficient Framework for Wideband Interference Mitigation in Automotive FMCW Radar Systems

Abstract

Millimeter-wave (mmWave) frequency-modulated continuous-wave (FMCW) radar is currently widely deployed in modern vehicles for advanced driver-assistance systems, and is regarded as one of the most promising sensing modalities for future autonomous vehicle systems. Compared with other major vehicular sensors, such as cameras and light detection and ranging, mmWave radar offers superior performance in terms of sensing range, velocity resolution, cost-efficiency, and robustness under low-visibility and adverse weather conditions. Despite these advantages, the sensing accuracy of FMCW radar can be severely degraded by inter-radar interference in high-density traffic scenarios, posing a significant obstacle to realizing Level-5 fully autonomous driving. To address this critical challenge, deep learning-based radar interference mitigation methods have recently attracted increasing attention and demonstrated promising performance. However, most existing approaches suffer from training-deployment mismatch and high computational latency. In this paper, we propose CMxUNet, a structurally enhanced UNet framework that embeds ConvMixer-based spatial-channel decoupled processing to effectively model globally coherent interference while preserving fine-scale target features. The proposed CMxUNet can be trained offline using synthetic data and deployed online in real-world environments with low-latency inference. Its performance is comprehensively evaluated and compared with state-of-the-art methods using both simulated datasets and real-world radar measurements collected from extensive experiments. Furthermore, a CMxUNet prototype is implemented on a real 79 GHz MIMO FMCW radar platform to demonstrate its effectiveness and feasibility for real-time automotive radar processing.