Key Technologies For3D Visualization Of Large-scale Urban Building Models By GPU Acceleration

As the application fields of digital earth are expanding, people’s requirement ofthe amount of information expressed by all kinds of spatial data is increasing. Also,with the concept of Smart City proposed, the virtual city buildings model data hasbecome another kind of important spatial data. On the other hand, the powerfulparallel computing ability of modern graphics hardware has gradually been applied tomany kinds of graphics issues or non-graphics issues, and achieves the GPUacceleration of these procedures.According to the above research backgrounds, taking the prospective of digitalearth, based on the theories and technologies of GPU acceleration, this thesis aims atthe efficient3D visualization of large-scale urban building models. Here we discussthe accelerations of several visualization procedures for the models data under thenetwork environment by making use of parallel computing of GPU. These proceduresinclude the stage of data scheduling, level of detail, realistic display and so on. Theaccelerations improve the efficiencies of the original CPU methods of theseprocedures, and strengthen the supporting of digital earth system for urban buildingmodels. Throughout the whole thesis, the main contributions of our work are asfollows:1. We propose to build spatial index of large-scale urban building models based onR tree and its variant to serve the scheduling of building models data. On thisbasis, we implement the acceleration of spatial query for R+tree by utilizingCUDA. Then, we propose a double speed-up algorithm for spatial query. Thisalgorithm maps the ranges of segmented view frustum to the thread blocks ofCUDA, and maps each entity to each block, reduces the running time of datascheduling greatly.2. On the basis of previous research work, we pre-store the hierarchical data of thesimplified building models into the CPU memory firstly. We propose a dynamicand efficient method for the transmission between CPU and GPU. Then wemodify the hierarchical data of building models transported in GPU video memory parallel by taking advantage of programmable shaders and then renderbuilding models of the corresponding level of detail in real time. By this way, wecan implement the GPU acceleration of LOD mechanism of large-scale urbanbuildings.3. By using the pixel shader of programmable pipeline, we transfer the parallelcomputing of Phong lighting model algorithm from vertex level of traditionalmethod to pixel level, so that the effects of shading of Phong lighting model canbe enhanced. We propose a GPU acceleration method for determination ofshadow body by making using of geometry shader. Also, we use graphicshardware to implement occlusion query to accelerate the judgment of whether torender the shadow volume generated by the specific scene node. Thus, therendering efficiency of shadow is improved.We verify the efficiency of GPU algorithms of each stage by experiments. Theexperiments show our methods are feasible and also efficient, can satisfy the3Dvisualization of large-scale building models of digital earth, and provide technicalsupports for the new generation of digital earth product.