Anti-Occlusion UAV Target Detection Based on Attention Feature Fusion

UAVs have developed rapidly in recent years, and their scope of application has been expanding. However, it has also given rise to various problems, such as illegal surveying and mapping and flights in no-fly zones. In order to quickly detect UAV targets during these events and take appropriate measures, it has become particularly important to study UAV target detection. For UAV images captured in real scenarios, there is object occlusion and background blending in UAV images. This makes it difficult to detect UAVs hidden in their surroundings. Therefore to solve the problem that UAVs are difficult to detect, an improved UAV detection algorithm YOLOv7_AUTD (Anti-occlusion UAV Target Detection) is proposed based on YOLOv7. The improvement includes: firstly, FPN network integration with the CSP module, ensuring lightweight while obtaining gradient flow information; secondly, adding CA attention mechanism, while introducing channel attention mechanism, taking into account spatial attention mechanism; finally, the Focal loss function is used to solve the problem of imbalanced proportion of positive and negative samples, and the E-iou function is used for more accurate boundary box regression to improve the accuracy of localization. The experimental results show that compared with the YOLO7 algorithm, the proposed algorithm improves AP by 2.72%.