Eagle-YOLO: An Eagle-Inspired YOLO for Object Detection in Unmanned Aerial Vehicles Scenarios

Object detection in images taken by unmanned aerial vehicles (UAVs) is drawing ever-increasing research interests. Due to the flexibility of UAVs, their shooting altitude often changes rapidly, which results in drastic changes in the scale size of the identified objects. Meanwhile, there are often many small objects obscured from each other in high-altitude photography, and the background of their captured images is also complex and variable. These problems lead to a colossal challenge with object detection in UAV aerial photography images. Inspired by the characteristics of eagles, we propose an Eagle-YOLO detection model to address the above issues. First, according to the structural characteristics of eagle eyes, we integrate the Large Kernel Attention Module (LKAM) to enable the model to find object areas that need to be focused on. Then, in response to the eagle's characteristic of experiencing dramatic changes in its field of view when swooping down to hunt at high altitudes, we introduce a large-sized feature map with rich information on small objects into the feature fusion network. The feature fusion network adopts a more reasonable weighted Bi-directional Feature Pyramid Network (Bi-FPN). Finally, inspired by the sharp features of eagle eyes, we propose an IoU loss named Eagle-IoU loss. Extensive experiments are performed on the VisDrone2021-DET dataset to compare it with the baseline model YOLOv5x. The experiments showed that Eagle-YOLO outperformed YOLOv5x by 2.86% and 4.23% in terms of the mAP and AP50, respectively, which demonstrates the effectiveness of Eagle-YOLO for object detection in UAV aerial image scenes.

Automated summary

This article introduces an Eagle-YOLO detection model based on the characteristics of eagle eyes to address the challenges of object detection in UAV aerial photography images. The model integrates the Large Kernel Attention Module and introduces a large-sized feature map with rich information, and adopts a weighted Bi-FPN network and Eagle-IoU loss to improve the performance of object detection in UAV aerial image scenes.