YOLO for Small Objects in Aerial Imagery: A Performance Evaluation

Unmanned aerial vehicles (UAVs) offer significant advantages in accessing remote or high-risk locations. Yet, they also present challenges concerning privacy and safety. Effective air traffic management is crucial for ensuring compliance with security regulations. The primary objective is to monitor the flight trajectories of target objects in the airspace, necessitating efficient object detection, especially for distant objects, to facilitate surveillance and preparedness for addressing potential threats. However, the current performance of existing models in detecting small objects is inadequate for real-time surveillance applications. This study advocates prioritizing object detection over precise prediction of object bounding box sizes. We propose a novel approach where the efficacy of object detection is measured by the ability to identify small objects. This research serves as an additional variable for evaluating and enhancing models designed for small object detection. We selected several versions of the YOLO model and assessed them on the COCO val 2017 dataset, counting the number of accurately predicted objects. Based on the most popular drone size of 100–400 pixels, YOLOv5-x is best suited for the future development of smaller objects. However, because the model is large, we can choose the YOLOv5-s, which is the smaller version. YOLOv5-s detects small objects in the same way that the newest version of the YOLO family, the YOLOv9-e. By measuring this, we may establish the model's capabilities and select a model appropriate for the target object's size.