Deep Learning-based Object Searching and Reporting for Aerial Surveillance Systems

Multi-class aircraft recognition is important in aerial surveillance applications to make consistent proposals for decision makers. Motivated by the state-of-art object detection methods for aerial sensing images, we explored and achieved satisfactory results based on standard YOLO archi-tectures by providing an analysis of certain aerial scenarios, from chained to scattered or yin-yang pairs objects and model's performance comparison on a typical wild emergency situation. First, we tackle the scarcity of multi-class aircraft detection data by developing of a 50 cm GSD dataset. Then, applying different training strategies and using a series of good practices, we achieved a F1-score of 0.820 and a con-siderable mAP@50 of 0.809 for the more difficult detection class. Experiments have been conducted over a dataset from Google Earth platform. Thus, our final proposal for aerial surveillance systems contains the pre-trained YOLOv5×6 architecture with attention on performance maximization for multiple specific aerial scenarios, processing 50km2 in no more than 6 seconds.