AerialPerson-YOLO26

Single-class (person) detectors for people in aerial imagery from UAVs. They were trained on the four-source AerialPerson-Mix dataset for the IEEE SSRR 2026 paper One Operator, Many Robots: Natural-Language Command of UAV–UGV Teams for Search, Victim Detection, and Rescue. The deployed detector is AerialPerson-YOLO26l.

Full data, evaluations, scripts and the reproduction workflow are in the Zenodo archive: https://doi.org/10.5281/zenodo.20643213 (the concept DOI, which always resolves to the latest version).

Results

Precision–recall of the five sizes on the 797-image validation set, with maximum-F2, maximum-F1 and P≥0.90 cutoffs. The diamond marks YOLO26l at the 0.60 track gate.

YOLO26l trained on all four sources or on one source, evaluated on each validation slice and on pooled validation (Mixed). Outlined cells mark each specialist's own source. Mixed is pooled validation, not the average of the four sources.

Annotated flight video

Click the frame to open the 720p video (about 24 MB). The 97-second clip is from the May 2026 field trials. Settings: YOLO26l, inference size 640 px, showing only detections with confidence above 0.60. Results: 612 of 2,923 frames have at least one detection, 630 detections in total. The source is 4K, so each frame is downscaled to 640 px for inference and the smallest, most distant people may be missed.

Files

File	Contents
`AerialPerson-YOLO26l.pt`	Deployed detector (identical to `sweep/yolo26l/best.pt`)
`sweep/yolo26{n,s,m,l,x}/best.pt`	The five YOLO26 sizes trained on all four sources
`source_training_comparison/training/yolo26{n,s,m,l,x}/{heridal,ladd,nomad,sard}/best.pt`	20 single-source specialists

Validation results

Pooled validation set (797 images, 2,188 person boxes), mAP50–95, from re-evaluating the saved checkpoints:

Model	All four sources	Single-source specialists (range)
YOLO26n	0.412	0.093–0.176
YOLO26s	0.469	0.140–0.228
YOLO26m	0.487	0.131–0.236
YOLO26l (deployed)	0.502	0.071–0.235
YOLO26x	0.500	0.109–0.262

The deployed YOLO26l reaches mAP50 0.832 and mAP50–95 0.501 at its best training epoch. The validation split is a curated development set, not an independent field test, and NOMAD actors appear in both training and validation.

Usage

from ultralytics import YOLO

model = YOLO("AerialPerson-YOLO26l.pt")
results = model.predict("frame.jpg", conf=0.60)  # 0.60 is the confidence gate of the deployed system

The specialist sweep used Ultralytics 8.4.37.

License

Model weights: AGPL-3.0, consistent with the upstream Ultralytics YOLO framework. The AerialPerson-Mix labels and manifests are CC-BY-4.0. Source imagery is not included and keeps its providers' terms.

How to cite

If you use these data or models, please cite the paper and this record.

@inproceedings{jaraboPenas2026OneOperator,
  author    = {Jarabo-Pe{\~n}as, Alejandro and
               Bravo-Arrabal, Juan and
               Ser{\'o}n-Barba, Javier and
               Lin-Yang, Da-hui and
               Pastor, Francisco and
               Ladig, Robert and
               Fern{\'a}ndez-Lozano, Juan Jes{\'u}s and
               Garc{\'i}a-Cerezo, Alfonso and
               Christensen, Anders Lyhne},
  title     = {One Operator, Many Robots: Natural-Language Command of
               {UAV}--{UGV} Teams for Search, Victim Detection, and Rescue},
  booktitle = {2026 IEEE International Symposium on Safety, Security,
               and Rescue Robotics (SSRR)},
  year      = {2026},
  publisher = {IEEE},
  note      = {In press}
}

Record: https://doi.org/10.5281/zenodo.20643213 (concept DOI, always resolves to the latest version)

Downloads last month: 802

alejp1998
/

AerialPerson-YOLO26