【作者】 周圣川；

【导师】 陈戈；

【作者基本信息】 中国海洋大学 ， 计算机应用技术， 2014， 博士

【摘要】 城市是人类从事生产、生活的重要场所，城市场景的高真实感三维可视化仿真是构建数字城市、智慧城市的核心技术，并且已经渗透和广泛应用于虚拟地理信息系统、城市规划、智能交通等与人们日常活动息息相关的行业领域中。大规模城市三维可视化仿真涉及三维表示建模、海量数据处理、三维可视化渲染等理论、方法和关键技术的研究，需要综合考虑计算机的硬件与软件资源，实现大规模空间数据的建模存储、传输调度、分析计算以及可视化输出。本文通过分析城市三维可视化仿真的研究与应用现状，以数据组织管理和可视化方法为两个切入点，提出了现阶段需要解决的瓶颈与关键问题：第一，如何进行有效的场景建模，处理海量的城市三维数据，在保证模型视觉精度的同时尽可能地降低数据量，在CPU、GPU混合构架下实现数据的高效传输调度，为可视化系统提供数据支持；第二，如何实现大规模、城市级三维场景的视觉无损渲染，一方面保证模型的视觉精度与质量，另一方面对城市级大场景进行完整地展现，并且在性能上满足交互式乃至实时可视化的要求。针对以上问题，本文主要进行了四个方面工作。首先，从城市三维场景的构建入手，设计并实现了一种交互式的图像逆过程式建模（Inverse ProceduralModeling，IPM）方法和基于GPU的过程式重构算法。通过构造纹理图像的过程式语义表示并与几何建模进行无缝的融合，形成一种图形图像混合建模的虚拟城市三维表示方法，降低模型的空间复杂度，提高渲染性能。其次，在三维场景的渲染算法研究方面，针对城市三维场景的数据结构和视觉特征，提出一种基于混合图元的层次细节（Level-of-detail， LOD）模型。使用多重抖动和蓝噪声采样创建由点、线、多边形等离散图元构成的三维模型近似表示并编码存储，构造支持渐进LOD切换的多分辨率模型数据结构，并进一步地提出混合模型的渲染算法，实现高性能的城市级大场景可视化输出。然后，在可视化数据管理方面，通过设计面向虚拟城市的空间数据引擎，实现海量空间数据在外存、内存、显存之间的高效调度，优化大规模可视化系统的数据访问性能。最后，建立了完善的可视化算法视觉质量评价模型，综合了感知心理学、统计学、模式识别、计算机图形学等多个学科的理论，设计了结合主观的用户感知调查与客观的自动化算法的视觉质量评价方法，验证了本文构建的城市三维场景可视化系统的视觉无损性。本文的创新性工作和贡献主要有以下几点：1)在数据表示与建模方面，提出了一种逆过程式的城市三维场景图形图像混合建模与表示方法。通过提取城市建筑纹理图像中的重复和对称特征，构建基于语义的过程式语法，实现对纹理图像的建模与数据压缩。用户可以直接控制过程式语法的生成，并且可以快速、有效地去除原始图像中的行人、树木等遮挡噪声数据。在渲染算法方面，逆过程式建模的结果可以直接在GPU中进行过程式重构，无需引入额外的渲染循环或几何图元。与传统的三角面片加纹理的建模方式相比，该方法在保证模型几何和纹理视觉精度的情况下可以对建筑立面纹理实现约70%的数据压缩，大幅降低了城市级三维场景的数据空间复杂度，并且可以将模型的渲染性能提高20%左右。2)在渲染理论方面，提出了一种面向大规模城市场景的混合图元渲染方法。与原始模型相比，混合图元模型不需要纹理贴图，并且建立了模型复杂度与屏幕空间像素面积的直接关系，相对于原始模型和几何LOD模型分别实现了约10倍和4倍的渲染加速，可以实时渲染城市级大场景，并且在可视化效果方面明显优于传统的几何层次细节模型渲染方法。通过对可视化算法的视觉质量进行系统地建模与分析，提出了可视化算法的视觉质量评价方法，实现了视觉质量的指标化，完善了可视化算法的评价体系，使视觉质量有了定量和定性的评价方法。并且，进一步验证了本文渲染算法的视觉质量与原始模型相比无显著差异，是一种视觉无损渲染方法。3)在海量数据传输与调度方面，设计并实现了CPU、GPU异构体系下的空间数据引擎系统。与传统的空间数据引擎相比，通过Cell-Rtree空间索引结构和Hilbert空间填充曲线优化空间数据的检索和存储性能，可以实现高效的三维空间数据查询与数据访问。提出了适用于大规模可视化系统的外存、内存、显存数据交换策略，并针对三维可视化系统数据压力大、吞吐量要求高的特点，通过线程池、内存池、缓存算法等一系列策略优化了海量空间数据的传输效率，是实现大规模城市三维场景视觉无损渲染的数据服务基础。更多还原

【Abstract】 Urban simulation and visualization are the bases of digital city and intelligentcity systems, and also important to a variety of stakeholders such as virtualgeographics information system, urban planning, and intelligent transportation, etc.Rendering large-scale urban scenes requires an integration of3D city modeling, hugedata processing, and complex scenes rendering theories. The graphics processing unitneeds to work synchronously with the rest of the computer, and it requireshigh-performance access to a huge mount of spatial data thus provides high realisticrendering results through real-time computation and visualization.The key observation of our research is large-scale urban scenes simulationusually prone to huge data modeling and rendering issues. Firstly, rendering urbandataset requires a processing of large-scale geometric primitives and textures. How toreduce the amounts of data thus achieve the balance on visual quality and datacomplexity, and the implemention of efficient data storage, dispatching andtransmission on CPU-GPU heterogeneous system architecture are major bottlenecksin this field. Secondly, how to implement high performance urban visualization whichmaintains plausible visual quality and provides high realistic results is also an openproblem.The objective of this study is archieving visually lossless rendering of large-scaleurban scenes and resovling the critical issues in the field of urban visualization asmentioned above. Firstly, an inverse procedual modeling approach is introduced forurban textures modeling, and then seamlessly integrated with geometic models andprovides hybrid geometry-and image-based representation for3D urban modeling;secondly, a progressive hybrid level-of-detail representation which combines points-,lines-and splats-based models and created through multi-jittered sampling andrecursive Wang-Tiles is proposed; thirdly, an urban scenes oriented3D spatial dataengine is introduced for resolving the performance and throughout bottlenecksbetween hard drive, main memory and video memory; furthermore, a comprehensive visual quality accessment which combines cognitive psychology, statistics models,pattern recognition and computer graphics theories is established for evaluatingperceptual changes in the outputs of computer graphics applications. Theexperimental results confirmed that the visual quality of our urban visualizationsystem is indistinguishable from the original models.The main contributions of this study are as follow:1) A novel procedural representation for facade textures. It can be seemlesslyinteraged with geometries to create hybrid3D urban models. The user-assisted inverseprocedural modeling approach allows user to exploit repetations and symmetries offacades to create a splite grammar of the input and remove the occlusions in anefficient way. The procedural textures can be direcly reconstructed in GPU, so noextra rendering passes and geometries are introduced. Comparing with the geometiclevel-of-details, the hybrid of geometry and procedural facade modeling achieved acompression factor that averaged70%and improved the rendering performance by20%.2) A novel hybrid level-of-detail approach for large-scale urban scenesrendering. We combine point-, lines-and splat-based rendering to synthesizelarge-scale urban city images with high visual quality. The algorithm usesscreen-space projected area as a progressive LOD selector and saves large amounts ofmemory spaces by avoiding textures. Our implementation shows a10times speed-upas compared to the ground truth models rendered as full geometry, and is about4times faster compared to geometic LOD. The quality of the results is alsoindistinguishable from the originals as confirmed by a comprehensive visual qualityassessment metric for computer graphics applications. The assessment modelintegrates subjective perceptual user study and two objective algorithm-based metrics,thus implements evaluation and numeralization on human visual responses.3) Design and implementation of an urban oriented3D spatial data enginewhich is capable for working under CPU-GPU heterogeneous system architecture.High performance spatial querying is archived by the use of Cell-Rtree spatialindexing and Hilbert storage indexing algorithms.3D spatial data exchanging,dispatching and caching strategies, as well as threads pools and memory management models are also proposed for improving I/O performance in large-scale visualizationsystems. The spatial data engine could also be an important component for visuallylossless urban scenes rendering.更多还原