Vehicle Detection Method Based on Improved YOLOv8 in RGB-infrared Aerial Images

In recent years, RGB-infrared object detection in aerial images has received much attention. This is because there are complementary properties in the two modalities that can mitigate the adverse effects of single modality's own limitations. However, there are still challenges in detecting objects in low-light and occlusion conditions. The common occlusion in aerial images is non-object occlusion. Existing methods usually focus on the design of modal feature fusion strategies and do not specifically deal with the occlusion phenomenon. To solve this problem, a data augment method is proposed, which mixes the object with the surrounding background to simulate the actual occlusion situation, increases the proportion of occlusion samples in the dataset. In addition, for better feature capture, a cross-modal object detection network based on the improved YOLOv8 network is proposed. In the detection network, a multi-modal feature fusion module is designed, which can adaptively select channel features and suppress redundant features so as to fuse features from different sensor images more effectively. The method is evaluated on the DroneVehicle dataset, and experimental results show that our method achieves state-of-the-art performance.