M3-2.1 - Monocular Inter-UAV Distance Estimation Using Machine Learning and Geometric Scaling
- Event
- 23. ITG/GMA-Fachtagung Sensoren und Messsysteme 2026
2026-06-09 - 2026-06-10
Nürnberg - Band
- Vorträge
- Chapter
- Modellbildung und Signalverarbeitung 2
- Author(s)
- R. Pommeranz, K. Tebbe, J. Fischer, R. Heynicke, G. Scholl - Helmut-Schmidt-University, Hamburg
- Pages
- 450 - 455
- DOI
- 10.5162/sensoren2026/M3-2.1
- ISBN
- 978-3-910600-11-9
- Price
- free
Abstract
Reliable Inter-UAV distance measurement is a prerequisite for formation control and collision avoidance in cooperative Unmanned Aerial Vehicle (UAV) swarms. This paper presents the system integration and experimental characterization of a monocular distance estimation pipeline that combines a fine-tuned YOLOv₈s object detector with pinhole-model geometric scaling to convert bounding-box pixel widths into real-time depth estimates. The pipeline runs on an NVIDIA Jetson Orin NX companion computer via DeepStream SDK. The camera is intrinsically calibrated, and the effective projected width of each target UAV is determined from short-range hover reference measurements rather than nominal dimensions. An uncertainty propagation model is derived that combines camera calibration uncertainty, target width variability, and detector bounding-box jitter into a depth confidence interval. Two independent ground truth sources are used for validation: Laser Range Finder (LRF) measurements at thirteen waypoints (5 m – 50 m) with a DJI Matrice 4T target, and dual-rover RTK GNSS during a dynamic reference flight with a Holybro X₆₅₀ target. A mean absolute error (MAE) below 2 m is achieved for ranges up to 35 m, with the DJI Matrice 4T. The pipeline requires only a low-cost camera and runs on embedded hardware, making it directly applicable to vision-based swarm coordination on weightand power-constrained platforms.