New Parallel Decision Fusion Models for Improving Multiclass Object Detection in Drone Images

Automatic object detection in drone images is an important area of research due to its application such as surveillance and security, reconnaissance, search and rescue mission, land monitoring, self-navigating drones. It is considered a challenging task due to the distance, camera angle, and complex surroundings. Filtering predictions from object detection models and combining the results can help improve the performance of individual models. Selecting the best predictions from predictions of multiple models is called ensemble parallel decision fusion. We augmented the large-scale grand challenge VisDrone2019-DET image dataset to create VisDrone-split and VisDrone-overlap datasets. We trained seven level-0 large deep learning models on the drone datasets and propose four Parallel Decision Fusion (PDF) deep learning models that improve in terms of the multiclass object detection precision upon their corresponding level-0 models. The seven models were split into two sets of object detection models. The first set of level-0 models includes Faster R-CNN, RetinaNet, and YOLOv5s. The second set consisted of much larger level-0 models namely state of the art YOLOv5x, YOLOv7x, YOLOv8x, and EVA02L. All proposed PDF models improved upon their corresponding state-of-the-art level-0 models on VisDrone-split. The PDF model with the highest metrics was YOLOv8x-s + EVA02L-s, which scored an mAP50 of 0.601 and improved 2.39% upon the state-of-the-art level-0 EVA02L model.