A Lightweight Two-Step Detection Method for Real-Time Small UAV Detection

Uncrewed aerial vehicles (UAVs) have been widely adopted across various domains; however, their potential for malicious use cases, including espionage, illicit trafficking, and unauthorized surveillance, has escalated security threats. This growing risk underscores the urgent need for precise and timely detection systems that can ensure robust countermeasures against UAV-related threats. In response to this demand, various studies have explored deep learning-based UAV detection approaches that can attain increased detection accuracy and efficiency. However, several major challenges remain in this domain, including the difficulty of detecting small UAVs and the increased computational cost associated with achieving performance improvements for deep learning models. To address the frequent misclassification of small UAVs in one-stage object detectors, this study proposes a two-step detection method, as illustrated in the graphical abstract, which summarizes the proposed approach. First, YOLO object detection model identifies candidate small UAVs. Then, a Vision Transformer classifier refines these results by filtering out false positives. This integrated approach significantly improves detection accuracy for small UAVs that are easily confused with other flying objects. Furthermore, we propose a compression strategy that integrates pruning, weight initialization, and knowledge distillation (KD) to reduce the computational costs of real-time applications. Experimental results demonstrate that the proposed two-step method outperforms a diverse set of state-of-the-art object detection and classification models. Specifically, in experiments where multiple types of aerial vehicles are present, the proposed method achieves a 5.6% improvement in mean average precision (mAP) compared to conventional object detection models. Additionally, our compression approach preserves critical UAV features by selectively removing low-importance parameters, significantly reducing the degree of redundancy while minimizing the induced detection performance loss.