Research On Target Recognition And Localization Method Of UAV Aerial Images

The identification and positioning of targets has always been a hot spot in academic research and engineering applications.Aerial images obtained by small UAVs can be used to identify and locate specific types of targets in the images.It has great development prospects in agriculture,forestry,military affairs,and emergency rescue..This paper studies the method of target recognition and positioning using UAV aerial images and pre-stored prior satellite remote sensing images when the UAV cannot obtain real-time positioning information.At present,many related studies are based on complete image pair matching,which requires image segmentation,optimal search,template matching and other processes.There is no ability to search for the required target category first and then decide whether to perform matching and positioning.In response to such problems,this paper proposes a deep learning-based aerial image target recognition and positioning method.The main research work is as follows:(1)For small targets and overlapping occluded targets that are difficult to identify in aerial images,an improved YOLOv5 target recognition algorithm is proposed,including the construction of an additional small target detection layer,and the introduction of Swin-Transformer and Sim AM attention mechanisms.The improved model was verified on satellite remote sensing images and UAV aerial image datasets,which proved that the improved model improved the accuracy of target recognition.Then,the transfer learning between satellite remote sensing images and UAV aerial images is researched and applied,which improves the target recognition accuracy of UAV aerial images.(2)The commonly used traditional image matching methods are studied,and the improved ORB image matching algorithm is studied in depth according to the characteristics of aerial images and remote sensing images.support.(3)Based on the precise recognition of the specified category target and the basis of fast image matching realized above.A multi-strategy feature matching algorithm is proposed to significantly improve the matching accuracy of satellite remote sensing images and aerial images.And according to the prior ground target information in the satellite remote sensing image,the non-priori ground target position in the UAV aerial image is calculated.