An Adaptive Object Detection System for Aerial Surveillance Using Deep Learning
Open Access DepositedAccurate real-time object detection in aerial imagery is a challenging task for many deep learning models. Operational intelligence, surveillance and reconnaissance (ISR) systems leveraging such models struggle due to various factors, including limited training data availability, domain shift and concept drift, resulting in poor detection performance. Recently released object detection models aim to improve detection accuracy using specialized architectures, yet they often fail to advance in the aerial domain. Moreover, imagery is often resized during training to circumvent hardware limitations, reducing the features available for learning. To address these challenges, this praxis investigates the integration of an adaptive two-stage training pipeline into one of the latest You Only Look Once (YOLO)-based models, YOLOv12, using the VisDrone-DET dataset. Improvements include
95) at 51.96 frames per second (FPS).
• Two adaptive loss functions • Network improvements targeting small-scale object detection • An additional Complete Intersection over Union (CIoU) loss Results from the test set using 640-by-640-pixel imagery demonstrate a 36.7% mean average precision (mAP50
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.