Enhancing Object Detection in UAV Videos under Complex Environments
Keywords:
Unmanned Aerial Vehicle, RGB sensor, Selfcalibration, Color correction, YOLO11s, Precision–Recall, F1-score, mAPAbstract
The paper presents an improved computer vision system for unmanned aerial vehicles (UAVs) that combines RGB sensor lighting compensation with an optimized YOLO11s detection model. Unlike traditional digital correction methods (AWB, CLAHE, Histogram Equalization), the proposed approach is based on physical measurements of illumination and provides dynamic stabilization of detection confidence in real time. The use of white stripe padding during preprocessing eliminated geometric distortions and increased the stability of bounding box formation. Reducing the number of classes to four (armored_vehicle, support_vehicle, tank, vehicle) reduced the number of misclassifications and accelerated inference.
The YOLO11s model demonstrated high efficiency: the total mAP@0.5 is 0.939, and for the classes support_vehicle, vehicle, and tank, values of 0.983, 0.968, and 0.956 were achieved, respectively. Analysis of the Precision–Confidence, Recall–Confidence, and F1–Confidence curves confirmed the stable operation of the system over a wide range of thresholds, with a maximum F1 ≈ 0.89 at a threshold of about 0.4. The normalized confusion matrix shows a 20–35% reduction in false negatives. The model training time was reduced to 27 minutes, and the processing speed is 22–25 FPS with an additional power consumption of 0.18 W compared to the baseline UAV vision system without RGB sensor compensation.
Field experiments in conditions of grass cover, dry vegetation, smoke, and fog confirmed the effectiveness of sensory feedback and the stability of detection in complex visibility conditions.
References
AMSCO / AMS-OSRAM. (n.d.). TCS3472 / TCS34725 color light-to-digital converter datasheet. https://www.alldatasheet.com/datasheet-pdf/pdf/560511/AMSCO/TCS3472.html
AYAATech. (n.d.). Voltage specifications of a 3S LiPo battery explained. https://www.ayaatech.com/ru/news/voltage-specifications-of-a-3s-lipo-battery-explained/
Basova, O., Gladilin, S., Kokhan, V., Kharkevich, M., Sarycheva, A., Konovalenko, I., Chobanu, M., & Nikolaev, I. (2024). Evaluation of Color Difference Models for Wide Color Gamut and High Dynamic Range. Journal of Imaging, 10(12), Article 317. https://doi.org/10.3390/jimaging10120317
Cheng, D., Li, Y., Zhang, D., Wang, N., Gao, X., & Sun, J. (2022). Robust single image dehazing based on consistent and Contrast-Assisted Reconstruction. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence(pp. 848–854). https://doi.org/10.48550/arXiv.2203.15325
Espressif Systems. (2023). ESP32-S3 technical reference manual. https://documentation.espressif.com/esp32-s3_technical_reference_manual_en.pdf
Gui, J., Cong, X., Cao, Y., Ren, W., Zhang, J., Cao, J., & Tao, D. (2021). A comprehensive survey andtaxonomy on single image dehazing based on deep learning. arXiv. https://arxiv.org/abs/2106.03323
Hidayatullah, P., Syakrani, N., Sholahuddin, M. R., Gelar, T., & Tubagus, R. (2025). YOLOv8 to YOLO11: A comprehensive architecture in-depth comparative review. arXiv. https://doi.org/10.48550/arXiv.2501.13400
Liu, X., Yang, F., Wei, H., & Gao, M. (2022). Shadow compensation from UAV images based on texture-preserving local color transfer. Remote Sensing, 14(19), Article 4969. https://doi.org/10.3390/rs14194969
Murat, A. A., & Kiran, M. S. (2025). A comprehensive review on YOLO versions for object detection. Engineering Science and Technology, an International Journal, 70, Article 102161. https://doi.org/10.1016/j.jestch.2025.102161
NVIDIA. (2022). Jetson Nano developer kit documentation. https://developer.nvidia.com/embedded/jetson-nano-developer-kit
Raspberry Pi Foundation. (2023). Raspberry Pi High Quality Camera documentation. https://www.raspberrypi.com/documentation/accessories/camera.html
Sapkota, R., & Karkee, M. (2025). Ultralytics YOLO Evolution: An Overview of YOLO26, YOLO11, YOLOv8 and YOLOv5Object Detectors for Computer Vision and Pattern Recognition. ArXiv. https://doi.org/10.48550/arXiv.2510.09653
Shen, X., Cao, Y., Sui, B., Zhang, Sh., & Feng, D. (2025). An automatic remote sensing image shadow compensation method utilizing reflectance differences and transfer learning. GIScience & Remote Sensing, 62(1), Article 2487334.https://doi.org/10.1080/15481603.2025.2487334
Tang, G., Ni, J., Zhao, Y., Gu, Y., & Cao, W. (2024). A survey of object detection for UAVs based on deep learning. Remote Sensing, 16(1), Article 149.https://doi.org/10.3390/rs16010149
Ultralytics. (2023–2024). LetterBox and image resizing with aspect ratio preservation. https://docs.ultralytics.com/reference/data/augment/
Ultralytics. (2024). Ultralytics YOLO11 documentation. https://docs.ultralytics.com
Ultralytics. (2024–2025). Ultralytics repository: YOLO models (including YOLO11)[GitHub repository]. GitHub. https://github.com/ultralytics/ultralytics
Wu, X., Li, W., Hong, D., Tao, R., & Du, Q. (2021). Deep learning for UAV-based object detection and tracking: A survey. arXiv. https://doi.org/10.48550/arXiv.2110.12638
Zhao, B., Xu, Q., & Luo, M. R. (2020). Color difference evaluation for wide-color-gamut displays. Journal of the Optical Society of America A, 37(8), 1257–1265. https://doi.org/10.1364/JOSAA.394132
