High-Throughput UAV Video Processing: A Multithreaded Architecture for Real-Time Deep Learning-Based Image Analysis
Journal of Robotics and Control (JRC) · 23.11.2025 · 6 (6) · 2746-2759
DOI: 10.18196/jrc.v6i6.28366 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Robotics and Control (JRC), online erschienen am 23.11.2025. Verfasst von Thang M. Pham und 6 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Managing concurrent high‑resolution video from multiple UAVs is challenging because prior systems often optimize transmission or perception in isolation and rarely validate real‑time, multi‑stream operation with operator tooling and fault recovery. Research contributions are: a centralized, thread‑per‑stream architecture with a 3‑frame drop‑oldest capture buffer and on‑device post‑processing that sustains real‑time throughput while keeping the GUI responsive; standardized multi‑object‑tracking reporting with MOTChallenge metrics; quantified autonomous reconnection with low variance and disclosure of resource usage and reproducible engineering levers. Each RTSP feed is handled by a dedicated worker (capture→decode→infer→render), detection uses YOLOv8n with temporal identity assignment and inference is standardized (640×640, conf 0.5, NMS IoU 0.5). Experiments run on a laptop‑class platform with 720p streams. The system sustains 29.46 FPS (SD 0.50; 95% CI [29.36–29.56]) for a single stream and 28.56 FPS (SD 0.65; 95% CI [28.43–28.69]) for three concurrent streams (≈3.1% drop), with utilization increasing from CPU 25.86%→50.24% and GPU 21.73%→35.76% yet remaining below saturation. Recovery is 6.16 s after UAV reboot and 3.51 s after camera disconnect, without disrupting other feeds. Under MOTChallenge (IoU 0.5), daytime tracking attains MOTA 0.84, IDF1 0.87, MT 92% / ML 8%; partial occlusion yields MOTA 0.73, IDF1 0.70, MT 78% / ML 22%. These results demonstrate real-time, multi-stream operation with quantified robustness. The current scope is limited to one and three streams with a single YOLOv8n backbone, and these constraints motivate future scaling to five or more concurrent streams as well as comparative evaluations across YOLOv8 variants and higher-capacity YOLO models.
Zitieren
APA
Pham, T. M., Nguyen, N. Q., Hoang, M. V., Dang, V. A., Do, N., Truong, H. X., & Vu, N. D. (2025). High-Throughput UAV Video Processing: A Multithreaded Architecture for Real-Time Deep Learning-Based Image Analysis. Journal of Robotics and Control (JRC), 6(6), 2746-2759. https://doi.org/10.18196/jrc.v6i6.28366
Vancouver
Pham TM, Nguyen NQ, Hoang MV, Dang VA, Do N, Truong HX, et al. High-Throughput UAV Video Processing: A Multithreaded Architecture for Real-Time Deep Learning-Based Image Analysis. Journal of Robotics and Control (JRC). 2025;6(6):2746-2759. doi:10.18196/jrc.v6i6.28366
BibTeX
@article{Pham2025,
author = {Pham, Thang M. and Nguyen, Nha Q. and Hoang, Manh V. and Dang, Viet A. and Do, Nam and Truong, Hung X. and Vu, Nam D.},
title = {High-Throughput UAV Video Processing: A Multithreaded Architecture for Real-Time Deep Learning-Based Image Analysis},
journal = {Journal of Robotics and Control (JRC)},
year = {2025},
volume = {6},
number = {6},
pages = {2746--2759},
doi = {10.18196/jrc.v6i6.28366}
}
EUVIMED