Rotated Aerial Object Detection Based on Improved YOLOv5 Method

Abstract

This paper increases the accuracy of rotated aerial objects detection under certain practical application scenario by proposed enriching the data set method. In order to reduce computing time and improve the generalization ability of the model, two selections before and after the data augmentation are designed. Different sizes of data set are used to train and the corresponding results are compared to present the influence of size of data set on the training. The data set with more rotated images achieved 67% after 100 epochs of training on YOLOv5 network with the pre-trained model YOLOv5s.pt measured by mAP@.5:.95, higher than that of the original size of data set (65.6%) and small size of data (61.5%). The small data set with half of images in the original data set has the smallest value of mAP@.5:.95. Our method is proved to be efficient to solve the detection problem of the UAV when it is circling on the altitude of 100 m and shows the potential of data augmentation on aerial object detection.