LCSA-IDNet: A Lightweight Channel-Spatial Attention Network for Automotive FMCW Radar Interference Detection

With the growing density of automotive millimeter-wave radar deployments, mutual interference has become a critical issue, highlighting the need for effective interference detection and mitigation. Traditional interference detection methods can only detect the presence of interference but fail to precisely locate or quantify it, limiting subsequent suppression techniques. This paper proposes a lightweight channel-spatial attention network for interference detection (LCSA-IDNet), which leverages the distinct time-frequency characteristics of interference and target echoes via short-time Fourier transform (STFT). Different from the traditional detection methods, we define oriented bounding boxes (OBB) for precise interference detection, supporting subsequent suppression techniques, such as notched methods. LCSA-IDNet integrates the convolutional block attention module (CBAM) attention mechanism with the YOLO 11 baseline to enhance feature fusion and time-frequency domain continuity, enabling precise interference detection and localization. A hierarchical feature compression strategy is also utilized to reduce computational complexity while maintaining accuracy. We construct a new dataset to conduct experiments, demonstrating that LCSA-IDNet achieves an mAP50-95 of 84.9% in complex scenarios with a 16.3% improvement and 3.7% faster inference speed over YOLO 11. Additionally, preliminary experimental verification has been conducted on the AWR2243 platform.