Object Detection in UAV Ground-Based Visual Landing Process Based on Improved Faster R-CNN

Abstract

AbstractBased on the object detection of the UAV ground-based visual landing process, self-built aerial object dataset. The dataset was used for training and testing. In view of the fact that deep learning detection algorithm does not require manually designed features and has robust detection performance, this paper improves the deep learning based object detection Faster R-CNN and uses it for aerial object detection. Combining the characteristics and requirements of aerial object detection tasks, improvements to the anchor box settings are proposed. Increase 9 anchor boxes to 15 anchor boxes. The detection accuracy of the UAV landing process when the image changes greatly has been improved, and better results have been obtained. The experiments show that the mean accuracy of the test results on the dataset is improved by 4.62% compared to the mean value before the improvement.KeywordsUAV landingGround-based visionFaster R-CNNAnchor boxObject detection