Research On Feature Detection In Aerial Video And Its Application

With the population of the unmanned aerial vehicle in recent years,aerial photograph comes to people’s common life.Aerial photograph,namely air photography,is a photograph method which shoots the earth,buildings,scenes and cities in the air.Aerial videos can rep-resent geographical features and cities clearly,so those videos can be utilized in the fields of national defense,environmental protection,land investigation and natural disaster evaluation.Image feature extraction and matching are foundations and key technologies of aerial video analysis.We handle lots of researches about those technologies and propose two image stitch-ing methods.Based on above studies,we develop a target identification system,which has been applied on AVIC Cup-International UAV Innovation Grand Prix.The main work of this thesis is summarized as follows:(1)Firstly,we introduce the origin and development of aerial photograph,as well as the key technologies of the aerial video analysis,namely image feature extraction and matching.Summarize achievements about image registration technologies based on features acquired by researchers all over the world.The applications and development of image registration based on features on aerial videos are discussed in detail,which would provide valuable references to future research works.(2)Analyze and summarize several common image registration methods based on fea-tures.Discuss in detail about those methods from two aspects,feature extraction and feature matching.Each registration method mentioned in this paper has been achieved.Advantages,disadvantages and application fields have been analyzed and summarized for those image reg-istration methods in practice.(3)An aerial video mosaic method based on reference frame exchanging and panorama size self-adaption is proposed in this thesis.Based on the innovation of image registration and image fusion via image registration based on reference frame exchanging and image fusion based on panorama size self-adaption,our method reduces the accumulative errors generated in sequence stitching and allocates memory reasonably.It achieves stitching result improvement.(4)A diverse scene hybrid stitching method based on graph is proposed in the thesis.Based on the innovation of image classification via sequential grouping,cross group retrieval and global stitching,our method can identify diverse scenes and save time wasted in image reg-istration which are problems in sequence stitching.It achieves scene identification and panora-ma generation in the aerial videos which are temporal continuity and spatial repeatability.(5)Build an onboard-based ground drone identification system.Based on the innova-tion of devices selection and identification methods,we solve the long delay and image shake problems generated in wireless image transmission.It achieves 60 frames per second when capturing,95%recognition rate and 20 frames per second while processing.This system is successfully applied on Third term AVIC Cup-International UAV Innovation Grand Prix.