AIAA SciTech 2026 · Vision-Based DAA
Detect-and-avoid (DAA) is the safety-critical layer underneath every routing and airspace concept on this site: an eVTOL flying a flow corridor still has to see and avoid other traffic in real time. This project is a comparative study of three object-detection architectures, DETR, YOLOv9, and YOLOv11, for vision-based aircraft detection in low-altitude UAM airspace, trained and evaluated on 72,000+ images from the AVOIDDS dataset.
Three architectures, one accuracy-versus-speed trade-off: transformer-based detection tops out accuracy, while the YOLO family trades a few points of mAP for inference speed and a lighter compute footprint, the deciding factor for what actually fits on an eVTOL or UAS flight computer.
Best raw detection accuracy of the three, at the cost of the slowest inference: the reference point the two YOLO models are measured against.
~2.5x faster inference than DETR, for a 5.1-point mAP trade-off, the strongest accuracy-per-millisecond balance of the three.
Smallest compute footprint of the three, the candidate for compute-constrained eVTOL/UAS flight computers where DETR or YOLOv9 wouldn't fit.
Training and evaluation pipeline for all three architectures.
Pretrained DETR, YOLOv9, and YOLOv11 checkpoints.
Versioned, citable snapshot of the codebase on Zenodo.