Rotator-YOLOv5: Improved YOLOv5 for Vehicle and Vessel Detection in UAV Images

In recent years, the widespread application of UAVs in national defence, military, and other fields, coupled with rapid technological advancements, has led to closer integration of UAVs and computer vision technology. Aerial images captured by UAVs hold significant potential for data mining; however, these images often feature arbitrarily rotated targets and complex backgrounds, resulting in insufficient detection accuracy. To address this challenge, we propose the Rotator-YOLOv5 algorithm, an enhancement of YOLOv5.(1) We integrate Circular Smooth Label (CSL) into YOLOv5 to achieve angle prediction with reduced computational complexity. (2) We design a two-branch structure that combines Convolutional Networks and Transformers, enabling feature information extraction from local and global perspectives. (3) By incorporating Partial Convolution (PConv) and RepNCSPELAN4, we further enhance the speed of our algorithm. Experimental results demonstrate that Rotator-YOLOv5 achieves a mean Average Precision (mAP) of 44.5% and a mAP_0.5 of 80.9% on UAV aerial images with an input size of 1 024x1024, outperforming YOLOv5s by 1.1% and 6.1% respectively while maintaining real-time inference speed. Due to its lower deployment cost, Rotator-YOLOv5 is well-suited for real-time target detection tasks on embedded terminals with vertical viewpoints, such as robotic arms and UAVs.