大数据量场景的实时仿真既是进行数字地球应用技术研究的主要内容之一，也是地理信息系统向三维立体分析纵深发展所需解决的重要命题之一。随着三维场景数据量的日益增大以及专为图形渲染设计的GPU的普及，在不明显降低图形质量和复杂程度的前提下，研究并采用一些可以应用在通用计算机平台上的高效数据处理方法与图形绘制算法，以提高图形绘制速度，解决大数据量仿真场景在速度、质量及场景复杂度之间越来越突出的矛盾，正成为有重要意义的研究方向。
本论文主要依托于国家科技攻关项目“奥运环境遥感动态监测”和中国科学院知识创新科技奥运项目“奥运主场馆区工程环境高分辨率遥感监测与虚拟仿真研究”，进行了大数据量场景中模型的自动简化、视点相关的有选择性实时细化以及GPU支持下的优化处理等相关理论与算法等方面研究，完成了奥林匹克公园仿真系统三维场景网络发布与浏览的设计与开发等工作，实现了面向互联网的三维场景显示、模型操作和属性数据查询等功能。
本文研究工作与创新之处主要体现在：
1）研制出三维模型自动简化预处理模块，能有效降低人工简化编辑大数据量三维场景的作业强度。在二次误差度量算法的基础上，考虑边界约束条件和法向量限制，扩展了算法的应用范围。此种方法在尽量保持模型外观前提下，达到减少模型三角形数量，从而生成场景中不同细节层次模型的目标。同时能根据网络上发布三维场景的要求，缩减原始模型的文件尺寸到合适的大小。从三维成像效果上看，与普通二次型度量简化方法相比，改进算法得到的结果在外观显示上与原始模型能保持更高的相似性，有较好的显示质量。
2）实现了视觉相关的实时有选择性细化算法，进一步提高了场景渲染速度。应用递进网格表示方法高效的存储了网格信息，支持多细节层次场景的自动生成和光滑过渡，能获得快速可视化效果。为进一步加速渲染，实现了视点相关的有选择性格网，能显著提高目标区域细节，同时减少对最终图像贡献少或没有贡献区域的细节，有效降低了图形流水线的渲染负担。使用界定面概念为引入的递进格网有选择性细化方案提供了理论基础，列举和可视了所有可能的有选择性格网集合，从而在进行点分裂或边折叠进行时，不用顾及邻接区域的配置，不需要引入额外的顶点分裂或边折叠变换。在实现过程中，采用了受约束二次误差度量简化方法进行边折叠操作，运算速度快，并能够阻止面法向量的剧烈变化，防止三角形翻转现象发生，尽量保持了近似模型在外观显示上与原始模型特征的一致。
3）提出了一种利用上一帧三维场景成像信息指导本帧场景中结点选择和更新简化的方法。依据GPU具有的绘制到纹理功能和视觉成像理论，进行了上帧三维成像信息的保存和提取，并根据获取的颜色信息判断场景中目标的存在状况和遮挡关系，预先剔除了不必绘制和渲染的物体，从而加速了大数据量仿真场景的显示速度。
4）结合顶点缓存，生成了优化的几何数据布局。指出在几何数据处理过程中为避免CPU使用率达到上限，首先应该保证GPU足够的三角形吞吐量。其次，说明在使用FIFO顶点缓存进行最大程度上重复使用顶点时，将有利于提高模型格网的渲染速度。因此，引入了迭代切割与MLA两种方法以产生格网渲染顺序，生成的几何数据布局结果尽可能保留了格网局部性并保证了顶点重用，同时，配合了视点相关的实时有选择性细化方法，进一步提升了大数据量场景绘制的速度。
5）提出建立虚拟奥林匹克公园系统的基本框架，实现三维场景网络发布等功能。结合北京奥林匹克公园建设规划的具体情况，从功能模块、性能、结构组成等方面介绍了系统的设计与基于组件技术的开发流程，详细阐述了动态模型加入、数据库查询等技术问题的解决途径，最后完成了系统开发工作，实现了三维场景的网络发布与浏览、模型对象查询以及粒子系统特殊效果等功能。
关键词：奥林匹克公园 三维仿真 网络发布 大数据量场景 优化 GPU
Abstract
Massive data scene real-time simulation is not only the one of main researches for digital earth application techniques, but also the important problem to be resolved for geographical information system developing towards the 3D analysis phase. With the quantity of 3D scene data increasing and popularization of graphics processing unit specially designed for graph rendering, if not obviously decreasing the graph quantity and complex of 3D scene, the research that some effective data processing and graph rendering methods can be applied on general computers to speed up rendering and solve the conflict among the speed, quality and complex of massive data scene is becoming an important direction.
Supported by National Key Technologies R&D Program Olympic Games Dedicated Projects "Research of dynamic monitor for Beijing Olympic Green Environment" and the Chinese Academy of Sciences Knowledge Innovation Program Science and Technology for Olympic item – research of high-resolution remote sensing monitoring and 3D simulation for the Beijing Olympic area, the related theory and methods are discussed and applied in this dissertation including the automatic simplification for models in massive data 3D scene, the viewpoint relevant selective real-time refinement and optimally processing assisted by GPU. The design and development for internet release and web browsing of virtual Beijing Olympic Park system are accomplished. The functions such as the display of 3D scene, operating models and attribute information query are realized at last.
The main works and innovations of this dissertation are as follows:
a) The preprocessing module of automatic simplification for 3D models is accomplished. Based on the primary quadric error metric arithmetic, combined with boundary constrained condition and normal restriction, the primary arithmetic is modified and the application scope is extended. The new arithmetic can preserve the appearance of original model and decrease the quantity of triangles of models to build the level of detail for 3D scene, and keep down the file size to an appropriate one for internet release of 3D scene. From the point of 3D scene imaging, comparing to the primary arithmetic, the result obtained from the improved arithmetic can preserve higher similarity with original models and has better display quality.
b) The viewpoint relevant selective real-time refinement arithmetic is realized. The progressive mesh method is applied to effectively storage the mesh information and support the automatic building of multi LOD and smooth transition and the quick visualization effect can be obtained. In order to speed up the graph rendering, the viewpoint relevant selective mesh is implemented, which can obviously enhance the detail of target area and reduce the detail of some areas for less or no contribution to the ultimate imaging and effectively lessen the rendering burden of graph pipeline. The boundary piece concept is used to provide the theory base for the progressive mesh selective refinement scheme. All possible selective mesh sets are listed and visualized, which result in not introducing additional vertex split and edge collapse and not considering the configuration of neighbor area when implementing vertex split and edge collapse. In the procedure, the improved constrained quadric error metric arithmetic is adopted to accomplish the operation of edge collapse. The advantages are quick computing speed, preventing face normal dramatic changing, avoiding face flip and preserve the appearance of approximate model consistent to original model.
c) Based on the function of rendering to texture on GPU, a new node selection and update in 3D scene at this frame time by using the previous 3D scene imaging information. According to vision imaging theory and pBuffer technique, the previous 3D scene imaging information is stored and extracted at first, then models that have no contributing to last imaging are eliminated in advance by using color information to determine the existence status and occlusion among the models in 3D scene, thereby the display of massive data 3D scene is speeded up.
d) The idea that enough triangles throughput on GPU should be guaranteed to avoid the employing ration of CPU reaching the 100 percentages in the procedure of processing geometry data is pointed out at first. Then, because repeatedly using vertex data on FIFO buffer at the maximum extent will benefit to improve the mesh rendering speed, the two methods producing mesh rendering sequence, the recursive cutting and MLA, are introduced. The geometry data layout result preserves the local relationships and guarantees repeatedly using vertex. At the same time, this method can combine with the viewpoint relevant selective real-time refinement arithmetic to more speed up the massive data 3D scene rendering.
e) Combining with the plan of Beijing Olympic Park, the basic frame for constructing virtual Olympic Park is brought forward. The design of the system and developing flow based on components are discussed in detail from several aspects such as function module, performance and configuration. Then the relevant technique problems such as adding dynamic models into 3D scene and database query are explained. At last the system development is accomplished and some functions such as 3D scene internet release and web browsing, model attribute query and particle system special effect are realized.
Keyword: Olympic Park, 3D Simulation, Internet Release, Massive Data Scene, Optimization, GPU