Two‐Stage Approach to Small‐Object Detection

ABSTRACT Unmanned aerial systems (UAS) are increasingly finding applications in civilian and commercial sectors. The utilization of machine learning techniques in UAS image analysis significantly advances target detection and tracking algorithms. In the field of systems engineering, the integration of advanced object detection techniques within UAS represents a pivotal advancement. However, existing object detection and tracking systems encounter challenges when applied to aerial object detection, primarily due to the rapid changes and rotations of obstacles within the UAS's field of view during flight. This paper proposes a fast and accurate real‐time small object detection system based on a two‐stage architecture. Our solution addresses the challenges of small object detection by integrating traditional target detection with deep learning techniques. Specifically, it employs conventional background subtraction and deep learning algorithms to obtain initial detection boxes. Subsequently, we utilize target tracking techniques to refine and enhance the accuracy of the final detection results. By seamlessly integrating traditional and deep learning methods within a two‐stage architecture, our system effectively captures the dynamic nature of UAS flights, demonstrating improved accuracy and efficiency in small object detection. We evaluated our approach on small object datasets, and experimental results show that the proposed method enhances aerial object detection performance compared to conventional approaches. This research contributes to ongoing efforts to advance UAS applications across various domains. And by demonstrating the efficacy of our integrated approach, this research underscores the role of systems engineering in enhancing UAS capabilities.