AN EFFECTIVE APPROACH FOR LOW-ALTITUDE AERIAL OBJECT DETECTION

Nguyen Do Khai Hoan; Tran Do Thu Ha; Luu Hoang Minh; Nguyen Xuan Mong; Nguyen Van Dat; Truong Quoc Huy; Nguyen Thanh Binh

Nguyen Do Khai Hoan Hung Yen University of Technology and Education
Tran Do Thu Ha Hung Yen University of Technology and Education
Luu Hoang Minh Hung Yen University of Technology and Education
Nguyen Xuan Mong Hung Yen University of Technology and Education
Nguyen Van Dat Hung Yen University of Technology and Education
Truong Quoc Huy Hung Yen University of Technology and Education
Nguyen Thanh Binh Hung Yen University of Technology and Education

Keywords: Kite, object detection, UAV

Abstract

Over the past decade, advances in aerial surveillance systems have driven increasing interest in low-altitude flying object detection within the computer vision community. However, the effectiveness of detecting and classifying flying objects, such as unmanned aerial vehicles (UAVs) and kites, is strongly influenced by the availability of specialized datasets, which remain limited. To overcome this challenge, we introduce a new dataset named UKD (UAV and Kite Dataset). The UKD dataset includes 2,121 real-world images of UAVs and kites with bounding box annotations and is employed to train and assess four representative modern detection models, such as YOLOv7, YOLOv10, RT-DETRv3, and Salience-DERT. Experimental results demonstrate that leveraging the UKD dataset leads to notable performance improvements, with the detection results achieving up to 80.3%.