Unified radar-prior-guided multimodal learning for ground-based small UAV detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Detecting small unmanned aerial vehicles (UAVs) from ground-based sensors is challenging because distant targets occupy few pixels, provide weak visual cues, and are easily confused with complex backgrounds under illumination changes. Existing radar–vision detectors have incorporated radar cues at different representation and prediction stages; however, these interactions are typically designed for stage-specific objectives, while the use of a common target-location reference across representation and optimization stages remains insufficiently explored. We develop URPG-ML, a radar-prior-guided extension of RT-DETR that uses a shared image-plane radar prior as a common spatial reference from which stage-specific guidance is derived. Sparse radar observations are projected onto the image plane and spatially diffused to construct this prior; these operations establish spatial alignment and are not treated as the principal methodological contribution. The resulting prior is adapted to three distinct learning stages: radar-conditioned channel modulation for within-level representation, prior-conditioned adaptive scale aggregation for cross-level interaction, and reweighting of matched-target supervision. In this way, radar-derived spatial support is reused with consistent target-location semantics across multilevel representation learning and set-based detection optimization. On the public Multi-Modal Anti-UAV Dataset, URPG-ML achieves 95.4% mAP@0.5, 49.9% mAP@0.5:0.95, and 49.1% small-object average precision (APs). Compared with YOLOv11-dualBackbone, URPG-ML improves Recall, mAP@0.5:0.95, and APs by 6.1, 7.9, and 7.8 percentage points, respectively. Repeated ablation experiments over five random seeds further show that the principal performance trends are maintained across different training runs, while fine-grained and sensitivity analyses support the complementary roles of the stage-specific uses of the shared radar prior. These results demonstrate that stage-specific utilization of a common radar spatial reference provides an effective strategy for improving weak small-UAV representation, localization, and detection performance.