Deep Neural Network Architectures for Real-Time Drone Detection in Aerial Surveillance Systems: A Comparative Evaluation of YOLOv7, YOLOv8, and YOLOv11

The proliferation of consumer drones has raised security concerns for critical infrastructure, airports, and urban surveillance, creating a need for reliable real-time detection. This study presents a controlled comparative evaluation of selected YOLO generations — YOLOv7x, YOLOv8x, and YOLOv11x — for single-class drone detection under matched high-level settings: 640 by 640 input, 50 epochs, the AdamW optimizer at an initial learning rate of 0.001, and an effective optimizer batch size of 64. Experiments used the YOLO Drone detection Dataset (1,012 training and 347 validation images, 369 validation instances). All models were scored with one evaluation implementation applied to raw predictions, removing differences between framework-specific validators. Across five independent training runs, YOLOv8x achieved the highest mean average precision at an IoU threshold of 0.5, 0.912 with a standard deviation of 0.007, ahead of YOLOv11x at 0.892 (0.004) and YOLOv7x at 0.708 (0.008); a Welch t-test gives a p-value of 0.0012 for YOLOv8x against YOLOv11x. A provenance analysis found that 32 of the 347 validation images are near-duplicates of training images; re-scoring on the 315 remaining images lowers every result but preserves the ordering. Zero-shot evaluation on the independently collected DUT Anti-UAV test set, whose targets are predominantly small, reproduces the same ordering while absolute accuracy falls to 0.346 for the best model. Detection accuracy therefore does not increase monotonically with architectural generation, and accuracy measured on this dataset does not transfer to imagery dominated by small targets. YOLOv11x was the smallest model (56.8 million parameters) with the lowest peak inference memory, indicating a trade-off between accuracy and deployment cost.

@article{Abylay2026,
  title     = {Deep Neural Network Architectures for Real-Time Drone Detection in Aerial Surveillance Systems: A Comparative Evaluation of YOLOv7, YOLOv8, and YOLOv11},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Nazerke Abylay and Elvira Kadylbekkyzy and Bakhytzhan Kulambayev and Bayan Moldakalykova and Zharkyn Bimoldina and Saken Mambetov and Kuanysh Dossanbek},
  doi       = {10.14569/IJACSA.2026.0170932},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170932}
}