Research On Key Technologies For Constructing Three-dimensional Scenes For UAV Simulation Trainin

In recent years,unmanned aerial vehicle(UAV)technology has been rapidly developed in China and used in many fields such as urban management,agricultural farming,forestry survey,natural resource exploration,emergency rescue,film and television.At the same time,thousands of UAV operators urgently need to be trained for mastering theoretical knowledge,operation skills and application technology of UAV.To this end,the Ministry of Industry and Information Technology of the People’s Republic of China has released a report on promoting and managing the development of civil UAV industry in 2017,and proposed a series of supporting measures,including encouraging the qualified universities and vocational colleges to establish UAV-related majors,establishing UAV operator training and service system within multiple levels and various types,encouraging enterprises to recruit high-level UAV personnels from domestic and foreign countries,and strengthening the education and training of UAV operators.At present,nearly one hundred colleges and universities in China have established UAV-related majors,and a number of companies have specialized in UAV operator training,such as Shenzhen Dajiang Huifei,Shanghai Handong Haoxiang and Guangzhou Nengfei.Due to the particularity of UAV industry,UAV operator training has a series of challenges,including the shortage of teaching equipments due to the expensive prices of UAVs,frequent flight accidents and their potential danger to public safety and personals,and the contradiction between complex flight environments in real tasks and the limited and idealized training environments in schools.For the reason,it is important and meaningful to construct a virtual simulation training system,or VSTS for short,by employing the emerging technologies such as virtual reality,system simulation and artificial intelligence.Starting from the actual requirements of UAV simulation training and on the basis of the emerging technologies such as remote sensing,graphic processing and artificial intelligence,this thesis focuses on the key technologies,especially 3D scene modeling,for the construction of VSTS.The main content of the thesis and its innovations mainly include:First,we presented a simulation solution and related key technologies for VSTS on the basis of the analysis of the training objectives of UAV operators and some learning theories.UAVs have a wide range of applications within various complex outdoor scenes.One of the core objectives for UAV operators is mastering kinds of flying skills and be able to operate UAV to fly in various scenes skillfully and safely.To this end,VSTS needs to provide a variety of three-dimensional simulation training scenes for students and support them to train various flight skills repeatedly.In addition,from the perspectives of learning theories such as transfer learning,experiential learning and game-based learning,VSTS should provide more realistic training scenes,more natural interactive operating modes,and gamification training fashions.Aiming at the above obj ectives,this thesis proposes a method for constructing UAV simulation training scenes by using remote sensing data collected from real scenes in the real world.Compared with the traditional software-based manual modeling method,this method can construct three-dimensional digital models of various real scenes quickly and accurately.Secondly,we investigated the key technologies in the construction of three-dimensional scenes and proposed some novel methods,mainly including:(1)Based on the investigation of the acquisition technologies and preprocessing technologies of 3D scene data,a normal estimation method is proposed by detecting a group of maximum consistency neighborhoods,and the estimated normals are beneficial to preserve the complete shapes and sharp edges of objects in outdoor scenes;(2)Based on the investigation of existing 3D scene segmentation methods,a robust method is proposed for segmenting raw 3D point clouds of outdoor scenes meaningfully by generating supervoxels,extracting planes locally,and optimizing and mergering the extracted planes;(3)Based on the investigation of existing recognition technologies for typical objects in 3D scenes,a multi-scale convolutional neural network model is designed and trained for 3D point clouds of outdoor scenes,which is capable of identifying building roof,tree,road and so on by combining the learned deep features with a group of pre-estimated hand-crafted features with a high precision.Finally,aiming at the actual requirements for UAV simulation training,we investigated the methodology and pipeline for constructing 3D UAV virtual simulation training scenes,and presented two virtual simulation training modes.Given the images and point clouds of real scenes as well as a group of teaching aid models,a 3D UAV simulation training scene can be created flexibly and conveniently by using the proposed data processing methods and existing modeling softwares.At the same time,the flight process of UVA is simulated based on the real-time rendering technology of large-scale 3D scenes,and an interactive control method is presented by employing the real UAV remote controls.On this basis,simulation training can be carried out with two patterns:PC screen&remote control and VR glasses&remote control.