A Radio Frequency Sensor-Based UAV Detection and Identification System Using Improved Vision Transformer-Based Model

In this article, we propose a novel approach to addressing the challenge of unmanned aerial vehicle (UAV) detection and identification by combining object detection and image classification techniques. Diverging from prevailing research that predominantly relies on convolutional neural networks (CNNs), we introduce the radio frequency drone vision transformer (RFDroneViT), a novel two-stage vision transformer (ViT) consisting of two distinct sub-ViT. The first stage employs the designed end-to-end object detection with transformer (DETR) to detect the UAV, while the second stage utilizes ViT-B/16 to identify the detected UAV. Based on RFDroneViT, we propose several detailed optimization algorithms, culminating in the development of a cost-effective system capable of UAV detection and identification. We evaluate the system’s performance through rigorous experimentation. Our custom DETR model achieves the state-of-the-art performance in drone signal detection with an average precision (AP) of 74.3 and demonstrates an 8.7 improvement in small target AP (APS) for the signal detection task compared to the original DETR. Additionally, the classification model can achieve 98.7% Top-1. The experimental results demonstrated that our proposed system achieves great performance on drone detection and identification.