waleedraza93/aircraft-detection-paper

Aircraft Detection — DETR, YOLOv9, YOLOv11

Single-class (aircraft) object detectors trained and evaluated on the AVOIDDS vision-based aircraft detection dataset, developed for:

Raza, W., Stansbury, R. S., and Gharami, K. (2026). A Comparative Study of Vision-Based Detect and Avoid for Urban Air Mobility. AIAA SciTech Forum, Orlando, FL. AIAA 2026-0465. https://doi.org/10.2514/6.2026-0465

Training/eval code: github.com/raza-waleed/aircraft-detection-paper

Abstract

Urban Air Mobility (UAM) aircraft require reliable methods to avoid collisions at low altitudes, where radar-based solutions may be limited or costly. This work explores a vision-based Detect and Avoid (DAA) method using three detection models — YOLOv9, YOLOv11, and Detection Transformer (DETR) — all trained on the AVOIDDS dataset (72,000+ images of intruder aircraft under varied times of day, weather, geographic regions, and aircraft types). Compared to YOLOv8, DETR achieves the highest accuracy at 94.2% mAP, YOLOv9 provides 2.5x faster processing speed at 89.1% mAP, and YOLOv11 achieves 88.5% mAP — suitable for eVTOL and UAS platforms with limited compute.

Results

Model	mAP	mAP@50-95	Precision	Recall	File
DETR	94.2%	—	0.974	0.963	detr_aircraft.pth (epoch 48/50)
YOLOv9	89.1%	0.633	0.914	0.847	yolov9_aircraft.pt
YOLOv11	88.5%	0.643	0.911	0.841	yolov11_aircraft.pt

mAP figures are as reported in the paper. YOLOv9 delivers ~2.5x faster inference than DETR, making it the better fit where processing speed is constrained (e.g. onboard eVTOL/UAS compute); YOLOv11 trades a small amount of accuracy for a lighter footprint. DETR's mAP is not computed with the same IoU-sweep protocol as the YOLO mAP@50-95 column (hence the —). All three models are evaluated on the same held-out AVOIDDS validation split.

Usage

YOLOv9 / YOLOv11 (Ultralytics)

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

path = hf_hub_download("waleedraza93/aircraft-detection-paper", "yolov11_aircraft.pt")
model = YOLO(path)
results = model("image.jpg")

DETR (Transformers)

import torch
from huggingface_hub import hf_hub_download
from transformers import DetrImageProcessor, DetrForObjectDetection

path = hf_hub_download("waleedraza93/aircraft-detection-paper", "detr_aircraft.pth")
processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
model = DetrForObjectDetection.from_pretrained(
    "facebook/detr-resnet-50", num_labels=1, ignore_mismatched_sizes=True
)
model.load_state_dict(torch.load(path, map_location="cpu"))
model.eval()

All three models were trained on a single class: {0: "aircraft"}.

Dataset

Trained and evaluated on AVOIDDS — 72,000 labeled images of intruder aircraft under varied lighting, weather, geometry, and location.

• Dataset: Smyers, E., Katz, S., Corso, A., and Kochenderfer, M. (2023). AVOIDDS: A dataset for vision-based aircraft detection. Stanford Digital Repository. https://doi.org/10.25740/hj293cv5980
• License: CC BY 4.0
• Benchmark / simulator reference: sisl/VisionBasedAircraftDAA

The dataset is cited here, not rehosted.

Citing this work

@inproceedings{raza2026comparative,
  title     = {A Comparative Study of Vision-Based Detect and Avoid for Urban Air Mobility},
  author    = {Raza, Waleed and Stansbury, Richard S. and Gharami, Kanchon},
  booktitle = {AIAA SciTech Forum},
  address   = {Orlando, FL},
  year      = {2026},
  note      = {AIAA 2026-0465},
  doi       = {10.2514/6.2026-0465}
}

Code and this weights release are archived on Zenodo: 10.5281/zenodo.21818474

@software{raza2026aircraftcode,
  author    = {Raza, Waleed},
  title     = {raza-waleed/aircraft-detection-paper: v1.0.0 - AIAA 2026-0465 code and model release},
  year      = {2026},
  publisher = {Zenodo},
  version   = {v1.0.0},
  doi       = {10.5281/zenodo.21818474},
  url       = {https://doi.org/10.5281/zenodo.21818474}
}

License

• DETR weights (detr_aircraft.pth): Apache-2.0, inherited from the facebook/detr-resnet-50 base model
• YOLOv9 weights (yolov9_aircraft.pt) and YOLOv11 weights (yolov11_aircraft.pt): AGPL-3.0, inherited from the Ultralytics and YOLOv9 base implementations
• Dataset (AVOIDDS): CC BY 4.0 — cited above, not redistributed
• Training/inference code: MIT, see the GitHub repository