Particle-ellipse IoU loss for accurate drogue detection

Drogue detection is widely used in autonomous aerial refueling and drogue control, with high accuracy and speed requirements. However, since the image of the drogue is usually an ellipse with different axis ratios, the Intersection over Union (IoU) based on rectangular bounding boxes can hardly accurately measure the similarity between the predicted region and the ground truth. To solve the above issue, we propose a novel particle-ellipse IoU (PEIoU) loss for drogue detection in this work. By setting particles on the image and calculating the proportion occupied by the overlapping part, an estimation of true IoU for two elliptical regions can be obtained. Furthermore, an improved network based on YOLOv8 was designed, and the PEIoU loss was integrated into the model. Finally, a drogue dataset was created containing real images, which was used to perform experiments. The extensive experiments demonstrate that the method proposed in this paper not only outperforms the baseline but also operates efficiently in real-time on edge devices.