ARGUS-YOLO

ARGUS-YOLO — Human & Vehicle Detection from Nadir UAV Imagery

Three YOLO object detectors for detecting humans and vehicles (rescue forces, firefighters, emergency vehicles) in high-resolution nadir (top-down) UAV imagery of civil-protection and firefighting scenarios.

The models were developed for the ARGUS WebApp as part of the E-DRZ research project.

Classes: 0: human, 1: vehicle

Training

Each model was trained in two stages, starting from the official Ultralytics COCO-pretrained weights:

1. VisDrone — trained on the VisDrone-DET dataset (large public UAV benchmark, ~8.6k images) at 960 px input size. This establishes small-object detection capability on aerial imagery. VisDrone images are, however, mostly oblique views of urban traffic — not the target domain.
2. ARGUS fine-tuning — fine-tuned on our own ARGUS dataset at 1280 px input size: high-resolution nadir UAV captures of real firefighting and rescue operations and exercises (publication of the dataset is pending).

Intended Use & Inference

• Domain: nadir (top-down) UAV imagery of rescue / firefighting / civil-protection scenes
• Flight altitude: 20–100 m (as in the training data)
• Image resolution: high-resolution captures, ≥ 4000×3000 px (typical drone camera output)
• Input size: use imgsz=1280 at inference — the models were fine-tuned and evaluated at this resolution; other sizes will degrade results

from ultralytics import YOLO

model = YOLO("argus_yolo26x_1280.pt")  # or argus_yolo11l_1280.pt / argus_yolo11x_1280.pt
results = model.predict("uav_image.jpg", imgsz=1280, conf=0.25)
results[0].show()

Note: YOLO26 requires a recent ultralytics version (≥ 8.4). YOLO11 models work with any version that includes YOLO11.

Results

Evaluated on the ARGUS validation split (100 held-out images, 1389 annotations) with Ultralytics .val() at imgsz=1280, conf=0.001, iou=0.6, batch=1 (single-image deployment conditions; inference time on an RTX 5080).

Overall

Per class

Humans are the harder class: at 20–100 m altitude a person covers only ~43×44 px (median) even in 4000×3000 px images, which is why the models were trained on high input resolutions.

The ARGUS Dataset

The fine-tuning dataset consists of nadir UAV imagery from real firefighting/rescue operations and exercises in Germany (e.g. flood response in the Ahr valley 2021, fire exercises, DRZ integration sprints and oprations from the Bielefeld fire brigade). Publication of the dataset is pending.

• Resolutions: 2048×1534 up to 8000×6000 px; most common 4000×3000 (212 images) and 4056×3040 (113 images)
• Median object size (native resolution): human ≈ 43×44 px, vehicle ≈ 145×140 px

Sample images

Sample of holdout validation split — predictions of argus_yolo11x_1280 (imgsz=1280, conf=0.25, green boxes) vs. ground truth (red boxes):

Limitations

• Nadir bias. Fine-tuning data is almost exclusively top-down; performance on strongly oblique views relies on the VisDrone stage and will be weaker.
• Domain. Trained on European (German) rescue and firefighting scenes; generalization to other environments is untested.
• Resolution sensitivity. Results were obtained at imgsz=1280 on high-resolution inputs; low-resolution imagery or smaller inference sizes will reduce accuracy, especially for humans.

License

The models are derived from Ultralytics YOLO11 / YOLO26 pretrained weights and are therefore released under AGPL-3.0, matching the Ultralytics license.