R-UNet: A Robust UNet-Based Network for Millimeter-Wave Radar Target Detection

The safety and reliability of autonomous vehicles heavily depend on robust environmental perception technologies, especially under complex weather conditions such as rain, fog, and snow, where accurately identifying road targets becomes a critical challenge. Millimeter-wave radar, with its strong penetration capability in harsh environments, has emerged as a key perception technology for autonomous driving. Addressing the limitations of existing detection networks in fully utilizing both frequency-domain features and spatial context, we propose R-UNet, a highly robust pure millimeter-wave radar target detection network. R-UNet is built on the UNet architecture and incorporates a spatio-temporal mixing (STM) module designed for multi-frame and multi-chirp radar data. This module suppresses interference and enhances target responses through multi-branch frequency-domain feature fusion. A 3D token mixing module is also introduced at the bottleneck to improve spatio-temporal interaction. Experiments on the CRUW dataset show that R-UNet has outstanding performance in target detection in complex road conditions, providing a practical solution for pure millimeter-wave radar perception.