Deep Radar Perception Network With RTK Supervision for Slow-Moving UAV Detection

Reliable detection of slow-moving unmanned aerial vehicles under low signal-to-noise ratio conditions remains a core challenge for next-generation intelligent sensing systems. We propose AirSentinel, a novel deep radar perception network to leverage frame-synchronized real-time kinematic-derived geocoordinates for supervised learning, bridging the gap between image-based and radar-based methods. To address the challenge of weak feature discrimination and energy leakage in radar data, we propose a local search and guard band mechanism that identifies Doppler peaks across coherent processing intervals. We further propose an information weight calibration adapter and a global temporal modeling head to facilitate feature alignment across views and capture long-range temporal dependencies, respectively. To mitigate data scarcity, a object transformation and generation scheme is introduced, significantly boosting data diversity and annotation efficiency. Extensive experiments on the newly curated radarUAV-24K dataset demonstrate that AirSentinel outperforms state-of-the-art methods in both accuracy and real-time performance. This work establishes a robust paradigm for radar-based object detection, with implications for autonomous monitoring and urban airspace security.