Transformers-Based UAV Networking Approach for Autonomous Aerial Monitoring

The rapid deployment of Unmanned Aerial Vehicles (UAVs) across various industries, including surveillance, disaster management, and logistics, necessitates precise and efficient detection and tracking mechanisms. Traditional deep learning models, such as convolutional neural networks (CNNs), face limitations in handling small object detection, occlusion, and dynamic environmental conditions. Transformers, leveraging selfattention mechanisms, provide a promising alternative by capturing long-range dependencies and contextual relationships effectively. Hence, this paper proposes a transformer-based framework for UAV detection and tracking, incorporating advanced feature extraction techniques and self-attention mechanisms. The proposed model processes image patches, extracting spatially and contextually rich representations to enhance detection accuracy. A curated dataset of 7,620 images, covering diverse UAV scenarios, is employed for training and evaluation. The transformerbased model achieves a state-of-the-art mAP of 0.9981, outperforming established models such as YOLOv5 and YOLOX in terms of detection accuracy. The framework demonstrates strong real-time processing capabilities with an inference time of 31.627 ms and a frame rate of 37.44 FPS, ensuring seamless UAV monitoring. Evaluation metrics, including accuracy, precision, recall, and $\mathbf{F 1}$-score, further validate the model’s reliability in real-world scenarios.