J. Eur. Opt. Society-Rapid Publ. 22, 49( 2026) 489
Fig. 7. Qualitative detection results of different YOLO-based detectors under diverse scene conditions. From top to bottom, the scenes correspond to Woodland, Cloudy Mountain, Clear Sky, Urban Street, Rural Landscape, and Snowy Mountain. From left to right, the detection results are generated by YOLOX, YOLOv13, YOLOv12, and YOLOv6, respectively. Red dashed boxes highlight challenging regions, while blue bounding boxes indicate detected UAV targets.
Table 3. Summarizes the quantitative detection performance and model complexity of all six detectors, including the four YOLO-based detectors as well as Faster R-CNN and RT-DETR, on the proposed synthetic UAV swarm dataset.
Detector |
mAP 50 |
mAP 50 – 95 |
P |
R |
Params |
Model Size |
FPS |
YOLOX |
0.886 |
0.875 |
0.912 |
0.573 |
9.0 |
17.3 |
95 |
YOLOv13 |
0.903 |
0.897 |
0.934 |
0.612 |
9.0 |
18.5 |
75 |
YOLOv12 |
0.871 |
0.852 |
0.887 |
0.542 |
9.4 |
18.0 |
80 |
YOLOv6 |
0.879 |
0.868 |
0.901 |
0.558 |
18.5 |
38 |
105 |
Faster R-CNN |
0.842 |
0.813 |
0.856 |
0.487 |
41.1 |
160 |
22 |
RT-DETR |
0.862 |
0.847 |
0.889 |
0.531 |
32.0 |
64 |
48 |
In cluttered environments such as Woodland and Urban Street, false positives occasionally occur due to background elements such as tree branches or building edges exhibiting visual patterns similar to UAVs. This highlights the importance of background diversity when evaluating detection performance.
Despite the overall satisfactory detection performance, several failure cases are observed across all detectors. The most common failure cases occur when UAV targets are extremely small or partially occluded, particularly in the Clear Sky and Snowy Mountain scenes. In such cases, the lack of discriminative texture information leads to missed detections.