Deep Learning-Based Flying Object Detection using Avian-Airborne Database: Military Applications

Abstract

Due to the upsurge in terrorist attacks, aerial surveillance has become essential in numerous locations, including sports stadiums, airports, VIP residences, and border security. Real-time detection of flying objects such as aircraft, birds, and unmanned aerial vehicles (UAVs) is a very challenging task due to their small size, environmental background, relative viewing angles, and lightning conditions. Deep learning algorithms have recently acquired prominence in object detection compared to traditional approaches; however, the research towards flying object detection is still in the early stages due to the lack of adequate datasets. In addition, there are very few publicly accessible datasets in combination with various flying objects such as aircraft, birds, or UAVs. In order to overcome these issues, this paper introduces a new dataset named Avian-Airborne, which contains 2550 images of three categories of flying objects: aircraft, birds, and UAVs, with 850 images collected from each category. Further, the Avian-Airborne dataset and the existing Det-Fly dataset are evaluated using two prominent deep-learning algorithms, YOLOv5s and YOLOv8. By implementing the YOLOv5s technique, the mean average precision (mAP) for flying object detection on the Det-Fly and proposed datasets is 97.2% and 99.2%, respectively. Similarly, the corresponding values for the YOLOv8 algorithm are 99.5% and 99.6%, respectively. According to the evaluation results, it is observed that both the YOLOv5s and YOLOv8 algorithms show prominent results on the newly introduced Avian-Airborne dataset.