Shallow Feature Enhanced Regression Model for UAV Traffic Object Detection

The advancement of intelligent transportation systems has brought about the widespread implementation of unmanned arial vehicles. However, these vehicles face several challenges when it comes to object detection, such as low adaptivity to deformed objects, inaccurate detection of small and blurred objects, and poor dataset annotation quality. To address these issues, this paper proposes the S-UAV object detection method, which is based on a regression model and shallow feature enhancement. Specifically, we propose the DC2 module to optimize feature extraction, the SPAN network to strengthen shallow feature fusion, and the Wise-IoU method to optimize bounding box loss. Our results demonstrate that the mAP of the S-UAV model can reach up to 0.447, representing a 19.5% improvement over the YOLOv8. This method is of great significance for enhancing the application of UAV technology in ITS.