부산대학교 도서관

In this paper, we introduce a state-of-the-art deep learning technique designed to accurately differentiate between drones and birds. This technique is particularly effective in reducing hazards associated with unauthorized drones, especially in airport environments where such drones can cause significant flight disruptions. Our approach involves the utilization of a meticulously compiled image dataset for testing, yielding results that surpass previous detection methods outlined in existing literature. Among the models evaluated, ResNet18 emerges as a standout, achieving an impressive average precision (AP) of 0.739 in medium area ratios. A key feature of our method is its ability not only to detect drones but also to precisely distinguish them from birds. The dataset employed in this research is derived from the publicly accessible real-world data of the 2020 Drone vs. Bird Detection Challenge.