Abstract:

Accurate and rapid recognition of aerial targets is crucial for improving air traffic efficiency and ensuring flight safety. Currently， in the field of aviation flight data， most datasets for aircraft target detection tasks focus on ground-based stationary aircraft targets captured from a top-down view in remote sensing images. There is a lack of datasets specifically designed for the detection of dynamic aerial aircraft targets， which makes it difficult to support the theoretical research on algorithms and the practical application of models for dynamic aerial aircraft target detection tasks. To address this gap， this study collects and constructs the first Aerial Aircraft Detection Dataset （AADD） that includes annotations of both Vertical Bounding Boxes （VBBs） and Rotated Bounding Boxes （RBBs） for aerial aircraft targets. This dataset is compiled from screenshots of public videos， covering 13 categories， 4 121 images， and 4 409 aircraft instances. In this study， an image quality classification model is trained to realize the automatic screening and acquisition of high-quality screenshots from videos； subsequently， manual screening is conducted to further select images that meet quality requirements， and finally， manual annotation is used to provide AABB and RBB annotations for aircraft targets. In addition， benchmark testing is performed on common object detection algorithms， among which the YOLOv8 detection algorithm achieves the best performance with an mAP50-95 of 80.5%. This result demonstrates the high challenge of detection tasks on this dataset， while the test outcomes can serve as a performance benchmark to support related research by scholars.

Key words: aircraft, aerial target detection, public dataset, deep learning, rotated bounding boxes