Research on Aerial Photography Target Detection Methods

Aerial target detection is a detection technology that uses an aerial platform equipped with visual sensors and deep learning technology to autonomously identify and locate ground targets. This technology has been widely used in fields such as environmental monitoring, road traffic management, safety inspections, and epidemic control. In recent years, one-stage target detection technology has shown excellent speed advantages and has made significant progress in real-time detection. In that sense, the YOLO (You Only Look Once) series become the most classic single-stage object detector, have developed to the eighth version, which is called "YOLOv8". In our study, YOLOv8 is used as the baseline model to solve the challenges from aerial images. To make the YOLOv8 model better adapt to the current scenario, we introduce EIoU_Loss and SoftNMS mechanisms. The introduction of these technologies improves the accuracy of YOLOv8 in target detection. Our study found that EIOU_Loss can reflect the difference between target scale and confidence, so it has a better performance in vehicle detection. The SoftNMS mechanism has better performance in dealing with occlusion problems. On the other hand, we also use Automatic Color Enhancement defogging technology to deal with foggy scenes during the detection process. We conducted rigorous experimental verification of the model enhancement and defogging methods on the UAVDT dataset and VisDrone dataset. The results show that the introduction of these technologies improves the performance while ensuring the real-time performance of the detection model.